# coding: utf8
import colorsys
from datetime import datetime
from pathlib import Path
from time import time
from typing import Any
import os
import cv2
import matplotlib.pyplot as plt
import numpy as np
import numpy.typing as npt
import pandas as pd
import yaml
from pytz import timezone
from scipy.interpolate import make_interp_spline
from scipy.ndimage import median_filter
from scipy.stats import gaussian_kde, iqr
from skimage import filters
from skimage.color import lab2lch, lab2rgb, rgb2lab
from sklearn.cluster import AgglomerativeClustering
from sklearn.mixture import GaussianMixture
import warnings
black_px = np.array([0.0, 0.0, 0.0])
BASE_DIR = Path(__file__).resolve().parent.parent
[docs]
def std4elem(x: npt.NDArray[np.float64]) -> npt.NDArray[np.float64]:
"""
Compute the element-wise standard deviation relative to the mean of
non-zero elements in the input array.
This function calculates the deviation of each element from the mean
of all non-zero elements in `x`, returning an array where elements
originally equal to zero remain zero.
Parameters
----------
x : numpy.ndarray
A 1D NumPy array of type float64. Elements with value zero are ignored
in the mean and standard deviation calculation.
Returns
-------
numpy.ndarray
An array of the same shape as `x` containing the standard deviation
for each element with respect to the mean of non-zero elements.
Zero entries remain zero.
Examples
--------
>>> import numpy as np
>>> from cie_utils import std4elem
>>> x = np.array([1.0, 2.0, 0.0, 4.0], dtype=np.float64)
>>> std4elem(x)
array([1.52752523, 0.15275252, 0., 1.37477271])
"""
x_mean = np.mean(x[x > 0])
n = len(x[x > 0])
c = []
for xi in x:
c.append(np.sqrt(np.power(xi - x_mean, 2) / (n - 1)) if xi > 0 else 0)
return np.array(c)
[docs]
def normalize_img(img: npt.NDArray[np.float64], rimg: npt.NDArray[np.float64]) -> npt.NDArray[np.float64]:
"""
Normalizes an image based on a reference image.
Pixel values of the input image are divided by the corresponding values
of the reference image. NaN values resulting from division by zero are
replaced by zero, and values are clipped to the range [0, 1].
Parameters
----------
img : npt.NDArray[np.float64]
The image to normalize. Must be a NumPy array of type float64.
rimg : npt.NDArray[np.float64]
The reference image for normalization. Must be a NumPy array of type float64.
Returns
-------
npt.NDArray[np.float64]
The normalized image, with pixel values in the range [0, 1].
"""
norm_img: npt.NDArray[np.float64] = np.zeros_like(img, dtype=np.float64)
for c in range(3):
norm_img[:, :, c] = np.divide(img[:, :, c], rimg[:, :, c], out=img[:, :, c], where=rimg[:, :, c] != 0)
norm_img[:, :, c] = np.nan_to_num(norm_img[:, :, c], nan=0)
norm_img = np.clip(norm_img, 0, 1)
return norm_img
[docs]
def min_max(img: npt.NDArray[np.float64]) -> npt.NDArray[np.float64]:
"""
Normalizes an image using min-max scaling.
This method independently normalizes each color channel of the image
to a range of [0, 1] based on the minimum and maximum values of each channel.
It is sensitive to outliers.
Parameters
----------
img : npt.NDArray[np.float64]
The input image as a NumPy array of type float64, with 3 channels.
Returns
-------
npt.NDArray[np.float64]
The normalized image in the range [0, 1].
"""
norm_img: npt.NDArray[np.float64] = np.zeros_like(img, dtype=np.float64)
norm_img[:, :, 0] = (img[:, :, 0] - np.amin(img[:, :, 0])) / (np.amax(img[:, :, 0] - np.amin(img[:, :, 0])))
norm_img[:, :, 1] = (img[:, :, 1] - np.amin(img[:, :, 1])) / (np.amax(img[:, :, 1] - np.amin(img[:, :, 1])))
norm_img[:, :, 2] = (img[:, :, 2] - np.amin(img[:, :, 2])) / (np.amax(img[:, :, 2] - np.amin(img[:, :, 2])))
return norm_img
[docs]
def sd_by_px(img: npt.NDArray[np.float64]) -> npt.NDArray[np.float64]:
"""
Calculates the standard deviation per RGB pixel.
For each pixel in the image, calculates the standard deviation of its three
color components (R, G, B).
Parameters
----------
img : npt.NDArray[np.float64]
The input image as a NumPy array of type float64, with 3 channels.
Returns
-------
npt.NDArray[np.float64]
A 2D array containing the standard deviation of the RGB values for each pixel.
"""
height, width, channels = img.shape
image_array = img.reshape(-1, channels)
std_dev_rgb: npt.NDArray[np.float64] = np.std(image_array, axis=1)
std_dev_rgb = np.reshape(std_dev_rgb, img.shape[:-1])
return std_dev_rgb
[docs]
def sd_by_elem(img: npt.NDArray[np.float64]) -> npt.NDArray[np.float64]:
"""
Calculates the standard deviation per color component (R, G, B) for each pixel.
For each color channel (red, green, blue), this function calculates the
standard deviation of pixels that are greater than zero. If a pixel
is zero, its individual standard deviation is considered zero.
Parameters
----------
img : npt.NDArray[np.float64]
The input image as a NumPy array of type float64, with 3 channels.
Returns
-------
npt.NDArray[np.float64]
An array with the same shape as the input image, where each element
represents the standard deviation of the corresponding color component.
"""
img_r = img[:, :, 0].flatten()
img_g = img[:, :, 1].flatten()
img_b = img[:, :, 2].flatten()
std_array: npt.NDArray[np.float64] = np.zeros_like(img)
std_array[:, :, 0] = np.reshape(std4elem(img_r), img.shape[:-1])
std_array[:, :, 1] = np.reshape(std4elem(img_g), img.shape[:-1])
std_array[:, :, 2] = np.reshape(std4elem(img_b), img.shape[:-1])
return std_array
[docs]
def rm_bg(img: npt.NDArray[np.float64], sd_val: float, sdimg: Any | None = None) -> npt.NDArray[np.float64]:
"""
Removes the background from a normalized image based on a standard deviation value.
Pixels whose value (or the value in `sdimg` if provided) is less than
or equal to `sd_val` are considered part of the background and are set to zero.
Parameters
----------
img : npt.NDArray[np.float64]
The normalized input image, of type float64.
sd_val : float
The standard deviation threshold value. Pixels with a standard
deviation below this value are considered background.
sdimg : Any | None, optional
An optional standard deviation image to use as a mask.
If None, the input image `img` is used. Defaults to None.
Returns
-------
npt.NDArray[np.float64]
The image without background, where background pixels have been set to zero
and values have been clipped to the range [0, 255].
"""
# mode = max(set(data), key=data.count)
mask = (sdimg if sdimg is not None else img) > sd_val
img_nobg: npt.NDArray[np.float64] = np.zeros_like(img)
img_nobg[mask] = img[mask]
img_nobg = np.clip(img_nobg, 0, 255)
return img_nobg
[docs]
def rm_bg2channel(
chn: npt.NDArray[np.float64],
sd_val: float,
sdimg: npt.NDArray[np.float64] | None = None,
) -> npt.NDArray[np.float64]:
"""
Removes the background from an image channel based on a standard deviation value.
Pixels whose value in the channel (or the value in `sdimg` if provided)
is less than or equal to `sd_val` are considered part of the background and are set to zero.
Parameters
----------
chn : npt.NDArray[np.float64]
The input image channel, of type float64.
sd_val : float
The standard deviation threshold value. Pixels with a standard
deviation below this value are considered background.
sdimg : npt.NDArray[np.float64] | None, optional
An optional standard deviation image to use as a mask.
If None, the input channel `chn` is used. Defaults to None.
Returns
-------
npt.NDArray[np.float64]
The image channel without background, where background pixels have been set to zero
and values have been clipped to the range [0, 1].
"""
mask = (sdimg if sdimg is not None else chn) > sd_val
chn_nobg: npt.NDArray[np.float64] = np.zeros_like(chn)
chn_nobg[mask] = chn[mask]
chn_nobg = np.clip(chn_nobg, 0, 1)
return chn_nobg
# TODO: verify utility
[docs]
def img_preparation(
img: npt.NDArray[np.float64], rimg: npt.NDArray[np.float64], sd_val: float
) -> list[npt.NDArray[np.float64]]:
"""
Prepares an image by applying normalization, background removal, and color space transformations.
The input image is normalized with a reference image, its background is
removed using a standard deviation threshold, and then it is converted
to CIELAB and CIELCh color spaces to extract their channels.
Parameters
----------
img : npt.NDArray[np.float64]
The input image in RGB format (float64 values).
rimg : npt.NDArray[np.float64]
The reference image for normalization in RGB format (float64 values).
sd_val : float
The standard deviation threshold value for background removal.
Returns
-------
list[npt.NDArray[np.float64]]
A list containing the L, a, b channels of the LAB image, and the
C (chroma) and H (hue) channels of the LCh image.
If the dimensions of `img` and `rimg` do not match, returns a list
containing only the original input image.
"""
if img.shape == rimg.shape:
img_norm = normalize_img(img, rimg)
img_sd = sd_by_px(img_norm)
img_nobg = rm_bg(img_norm, sd_val, img_sd)
lab_img = rgb2lab(img_nobg)
lch_img = lab2lch(lab_img)
return [
lab_img[:, :, 0],
lab_img[:, :, 1],
lab_img[:, :, 2],
lch_img[:, :, 1],
lch_img[:, :, 2],
]
return [img]
[docs]
def get_pdf(
img: npt.NDArray[np.float64],
) -> tuple[Any, Any, Any, int]:
"""
Calculates the Probability Density Function (PDF) for image data.
This function filters non-zero pixels, calculates the number of bins
using the Freedman-Diaconis rule, and estimates the PDF using
Gaussian Kernel Density Estimation (KDE).
Parameters
----------
img : npt.NDArray[np.float64]
The input image as a NumPy array.
Returns
-------
tuple[npt.NDArray[np.float64], npt.NDArray[np.float64], Any, int]
- data : npt.NDArray[np.float64]
The pixel data (non-zero values) of the image, rounded to one decimal place.
- rng : npt.NDArray[np.float64]
The range of values over which the density is calculated.
- density : Any
The density values of the estimated probability density function.
- bins : int
The number of bins calculated using the Freedman-Diaconis rule.
"""
data = np.around(img[img != 0], decimals=1)
if data.size == 0:
# Return safe defaults for empty data
return data, np.array([]), np.array([]), 1
# Calculate the Interquartile Range (IQR)
iqr_value = iqr(data, rng=(25, 75)) # IQR between the 25th and 75th percentile
# Calculate the number of bins using the Freedman-Diaconis rule
n = len(data)
bin_width = 2 * iqr_value / (n ** (1 / 3)) # Bin width
data_range = np.amax(data) - np.amin(data) # Range of the data
if bin_width == 0:
bins = max(1, int(np.ceil(np.sqrt(n)))) # fallback when IQR is 0 (uniform data)
else:
bins = max(1, int(np.ceil(data_range / bin_width))) # Number of bins
# print("#" * 30, "BINS", bins)
steps = (np.amax(data) + 1 - np.amin(data)) / bins
rng = np.arange(np.amin(data), np.amax(data) + 1, step=steps)
density = gaussian_kde(data)
density = density(rng)
return data, rng, density, bins
[docs]
def display_images(
imgs: list[npt.NDArray[np.float64]],
cols: int = 1,
fsize: tuple[int, int] = (10, 5),
cmaps: list[str] | None = None,
titles: list[str] | None = None,
save_imgs: bool = False,
output_dir: Path = Path.cwd(),
) -> None:
"""
Displays a list of images in a grid.
Parameters
----------
imgs : list[npt.NDArray[np.float64]]
A list of images to display.
cols : int, optional
The number of columns in the display grid. Defaults to 1.
fsize : tuple[int, int], optional
The figure size (width, height) in inches. Defaults to (10, 5).
cmaps : list[str] | None, optional
A list of colormaps to apply to each image. If None, no colormap is applied.
Defaults to None.
titles : list[str] | None, optional
A list of titles for each image. If None, no titles are displayed.
Defaults to None.
save_imgs : bool, optional
If True, saves the images to a file instead of displaying them.
The filename is based on the first title if available,
otherwise "test" is used. Defaults to False.
output_dir : Path, optional
The directory where the images will be saved if `save_imgs` is True.
Defaults to the current working directory.
"""
dt = datetime.now()
if len(imgs) < cols:
cols = len(imgs)
if cols < 1:
raise ValueError("Number of columns must be at least 1.")
rows = int(np.ceil(len(imgs) / cols))
fig, axes = plt.subplots(rows, cols, figsize=fsize)
# Asegurar formato uniforme para axes
axes = np.array(axes).reshape((rows, cols)) if rows * cols > 1 else np.array([axes])
k = 0
for i in range(rows):
for j in range(cols):
if k >= len(imgs):
axes[i, j].axis("off")
continue
axes[i, j].imshow(imgs[k], cmap=cmaps[k] if cmaps else None)
if titles:
axes[i, j].set_title(titles[k])
axes[i, j].axis("off")
k += 1
for ax in axes[len(imgs) :]:
ax.axis("off")
if save_imgs:
output_dir.mkdir(parents=True, exist_ok=True)
filename = f"display_images_{dt.timestamp():.5f}.png"
fig.savefig(str(output_dir / filename), dpi=600)
else:
plt.tight_layout()
plt.show()
plt.close(fig)
[docs]
def display_hist(
imgs: list[npt.NDArray[np.float64]],
cols: int,
fsize: tuple[int, int] = (10, 2),
titles: list[str] | None = None,
xlabel: list[str] = None,
ylabel: list[str] = None,
save_imgs: bool = False,
save_csv: bool = False,
output_dir: Path = Path.cwd(),
) -> None:
"""
Displays histograms and Probability Density Functions (PDFs) for a list of images.
Parameters
----------
imgs : list[npt.NDArray[np.float64]]
A list of images (or image channels) for which to generate histograms.
cols : int
The number of columns in the histogram display grid.
fsize : tuple[int, int], optional
The figure size (width, height) in inches. Defaults to (10, 2).
titles : list[str] | None, optional
A list of titles for each histogram. If None, no titles are displayed.
Defaults to None.
xlabel : list[str], optional
A list of labels for the x-axis of each histogram. Defaults to ["x"] for all.
ylabel : list[str], optional
A list of labels for the y-axis of each histogram. Defaults to ["y"] for all.
save_imgs : bool, optional
If True, saves the histograms to a file instead of displaying them.
The filename is based on the first title if available,
otherwise "test" is used. Defaults to False.
save_csv : bool, optional
If True, saves the density data (rng and density) of each histogram
to a CSV file. Defaults to False.
output_dir : Path, optional
The directory where the images will be saved if `save_imgs` is True.
Defaults to the current working directory.
"""
if not imgs:
raise ValueError("The image list is empty.")
if cols < 1:
raise ValueError("The number of columns must be at least 1.")
xlabel = xlabel or ["x"] * len(imgs)
ylabel = ylabel or ["y"] * len(imgs)
titles = titles or [""] * len(imgs)
rows = int(np.ceil(len(imgs) / cols))
fig, axes = plt.subplots(rows, cols, figsize=fsize)
axes = np.array(axes).flatten() if isinstance(axes, np.ndarray) else [axes]
def export_csv(index: int, rng: np.ndarray, density: np.ndarray):
filename = output_dir / f"density_{index}_{int(time() * 1000)}.csv"
df = pd.DataFrame(zip(rng, density), columns=["x", "pdf"])
df.to_csv(filename, index=False)
print(filename)
for idx, (img, ax) in enumerate(zip(imgs, axes)):
data, rng, density, bins = get_pdf(img)
ax.hist(data, bins=bins, density=True)
ax.plot(rng, density)
ax.set_xlabel(xlabel[idx])
ax.set_ylabel(ylabel[idx])
ax.set_title(titles[idx])
ax.axis("on")
if save_csv:
export_csv(idx, rng, density)
for ax in axes[len(imgs) :]:
ax.axis("off")
output_dir.mkdir(parents=True, exist_ok=True)
if save_imgs:
filename = f"histograms_{datetime.now().timestamp():.5f}.png"
fig.savefig(output_dir / filename, dpi=600)
else:
plt.tight_layout()
plt.show()
plt.close(fig)
[docs]
def display_plot(
imgs: list[npt.NDArray[np.float64]],
fsize: tuple[int, int] = (10, 2),
cmaps: list[str] | None = None,
titles: list[str] | None = None,
) -> None:
"""
Displays a combination of image, histogram, and pixel frequency scatter plot
for each image in the list.
For each image:
1. The original image is displayed.
2. A histogram and its Probability Density Function (PDF) are generated.
3. A scatter plot of pixel value frequency is created.
Parameters
----------
imgs : list[npt.NDArray[np.float64]]
A list of images (or image channels) to visualize.
fsize : tuple[int, int], optional
The figure size (width, height) in inches. Defaults to (10, 2).
cmaps : list[str] | None, optional
A list of colormaps to apply to each image. If None, no colormap is applied.
Defaults to None.
titles : list[str] | None, optional
A list of titles for each set of plots. If None, no titles are displayed.
Defaults to None.
"""
rows = len(imgs)
fig, axes = plt.subplots(rows, 3, figsize=fsize)
for i, img in enumerate(imgs):
pixel_num = img.shape[0] * img.shape[1]
pix_percentage = np.around(len(img[img != 0]) / pixel_num, 2) * 100
pix_label = f"#{pixel_num}\n{pix_percentage:.2f}%"
if rows == 1:
axes[0].imshow(img, cmap=cmaps[i] if cmaps else None)
axes[0].set_title(titles[i] if titles else None)
data, rng, density, bins = get_pdf(img)
axes[1].hist(data, bins=bins, density=True)
axes[1].plot(rng, density, label=pix_label)
axes[1].legend()
data_flat = sorted(set(data.reshape(-1)))
pixel_freq = np.array([[dts, np.count_nonzero(data == dts)] for dts in data_flat])
axes[2].scatter(pixel_freq[:, 0], pixel_freq[:, 1], marker=".")
else:
axes[i][0].imshow(img, cmap=cmaps[i] if cmaps else None)
axes[i][0].set_title(titles[i] if titles else None)
data, rng, density, bins = get_pdf(img)
axes[i][1].hist(data, bins=bins, density=True)
axes[i][1].plot(rng, density, label=pix_label)
axes[i][1].legend()
data_flat = sorted(set(data.reshape(-1)))
pixel_freq = np.array([[dts, np.count_nonzero(data == dts)] for dts in data_flat])
axes[i][2].scatter(pixel_freq[:, 0], pixel_freq[:, 1], marker=".")
plt.tight_layout()
plt.show()
[docs]
def display_2d_scatter_plot(
imgs: list[npt.NDArray[np.float64]],
cols: int,
fsize: tuple[int, int] = (10, 2),
titles: list[str] | None = None,
) -> list[npt.NDArray[np.float64]]:
"""
Displays 2D scatter plots of pixel frequency for a list of images.
Each scatter plot shows pixel values (x-axis) against their
frequency of occurrence (number of pixels with that value, y-axis). Only
non-zero pixel values are considered.
Parameters
----------
imgs : list[npt.NDArray[np.float64]]
A list of images (or image channels) for which to generate scatter plots.
cols : int
The number of columns in the plot display grid.
fsize : tuple[int, int], optional
The figure size (width, height) in inches. Defaults to (10, 2).
titles : list[str] | None, optional
A list of titles for each scatter plot. If None, no titles are displayed.
Defaults to None.
Returns
-------
list[npt.NDArray[np.float64]]
A list of NumPy arrays, where each array contains the frequency data
of pixels ([pixel_value, frequency]) for each input image.
"""
plot_data: list[npt.NDArray[np.float64]] = []
if len(imgs) < cols:
cols = len(imgs)
if cols < 1:
raise ValueError("Number of columns must be at least 1.")
rows = int(np.ceil(len(imgs) / cols))
fig, axes = plt.subplots(rows, cols, figsize=fsize)
if 1 in (rows, cols):
for j, img in enumerate(imgs):
data = np.around(img[img != 0], decimals=1)
data_flat = sorted(set(data.reshape(-1)))
pixel_freq = np.array([[dts, np.count_nonzero(data == dts)] for dts in data_flat])
axes[j].scatter(pixel_freq[:, 0], pixel_freq[:, 1], marker=".")
axes[j].set_title(titles[j] if titles else None)
plot_data.append(pixel_freq)
else:
k = 0
for i in range(rows):
for j in range(cols):
data = np.around(imgs[k][imgs[k] != 0], decimals=1)
data_flat = sorted(set(data.reshape(-1)))
pixel_freq = np.array([[dts, np.count_nonzero(data == dts)] for dts in data_flat])
axes[i][j].scatter(pixel_freq[:, 0], pixel_freq[:, 1], marker=".")
axes[i][j].set_title(titles[k] if titles else None)
plot_data.append(pixel_freq)
if k + 1 == len(imgs):
break
k += 1
plt.tight_layout()
plt.show()
return plot_data
[docs]
def display_3d_scatter_plot(
images: list[npt.NDArray[np.float64]],
colors: list[str] | None = None,
title: str | None = None,
) -> None:
"""
Creates a 3D scatter plot for classified data.
Each dataset in `images` is represented as a series of points
in a 3D space, useful for visualizing cluster distribution.
Parameters
----------
images : list[npt.NDArray[np.float64]]
A list of NumPy arrays, where each array contains the coordinates
(e.g., L, a, b from LAB) of the pixels in a cluster or group.
Each array is expected to have 3 columns.
colors : list[str] | None, optional
A list of color strings for each set of points. If None,
colors are automatically selected. Defaults to None.
title : str | None, optional
The title of the plot. Defaults to None.
"""
fig = plt.figure()
ax = fig.add_subplot(projection="3d")
images = [images] if not isinstance(images, list) else images
for i, img in enumerate(images):
try:
ax.scatter(img[0], img[1], img[2], c=colors[i], label=f"# px g{i}: {img.shape[1]}")
except IndexError:
# Consider more specific error handling if the index is out of range
breakpoint()
ax.set_xlabel("A")
ax.set_ylabel("B")
ax.set_zlabel("L")
ax.legend()
ax.set_title(title)
plt.show()
[docs]
def get_csv_data(
now: str,
data: npt.NDArray[np.float64],
filename: str,
folder: Path,
col_names: list[str] | None = None,
) -> None:
"""
Saves data to a CSV file within a specific folder structure.
The save path is `folder/csv/now/filename.csv`.
Parameters
----------
now : str
A timestamp string to be used as a subfolder name within 'csv'.
data : npt.NDArray[np.float64]
The data to save, must be a 2D NumPy array.
filename : str
The name of the CSV file (without the .csv extension).
folder : Path
The Path object of the base folder where the 'csv/now/' structure will be created.
col_names : list[str] | None, optional
A list of names for the CSV columns. If None, no column names will be used.
Defaults to None.
"""
if col_names is None:
col_names = []
df = pd.DataFrame(data, columns=col_names)
filepath = folder / "csv" / now
filepath.mkdir(parents=True, exist_ok=True)
csv_name = str(filepath / filename) + ".csv"
df.to_csv(csv_name, index=False)
[docs]
def get_csv_data_from_images(
imgs: list[Any],
folder: Path,
col_names: list[str] | None = None,
timezone=timezone("America/Bogota"),
) -> None:
"""
Extracts pixel frequency data from a list of images and saves them to CSV files.
For each image in `imgs`, pixel frequencies are calculated
(pixel values vs. number of occurrences) and saved to a CSV file.
Files are organized by date and time within the specified `folder`.
Parameters
----------
imgs : list[Any]
A list of tuples where each tuple contains (image_name: str, image: npt.NDArray[np.float64]).
The image must be a NumPy array. Pixels with zero value are ignored.
folder : Path
The Path object of the base folder where the CSV file structure will be created.
col_names : list[str] | None, optional
A list of names for the CSV columns (e.g., ["pixel_value", "frequency"]).
If None, no column names will be used. Defaults to None.
"""
if col_names is None:
col_names = []
now = datetime.now(timezone).strftime("%Y-%m-%d_%H%M%S")
for name, img in imgs:
data = np.around(img[img != 0], decimals=1)
data_flat = sorted(set(data.reshape(-1)))
pixel_freq = np.array([[n, np.count_nonzero(data == n)] for n in data_flat])
get_csv_data(now, pixel_freq, name, folder, col_names)
[docs]
def extract_segmentation_main() -> None:
"""
Main function for image segmentation extraction.
Prompts the user for the necessary paths for the image, reference image,
label file, and destination folder, then calls `extract_segmentation`.
"""
img_path = enter_str_input("Enter the absolute path of the image: ")
rimg_path = input("Enter the absolute path of the reference image (enter if not): ")
label_path = enter_str_input("Enter the absolute path of the image label: ")
dest_path = enter_str_input("Enter the absolute path of the destination folder: ")
dataset_yaml_path = enter_str_input("Enter the absolute path of the dataset.yaml file: ")
extract_segmentation(img_path, rimg_path, label_path, dest_path, dataset_yaml_path)
[docs]
def hsv_to_rgb(h: float, s: float, v: float) -> Any:
"""
Converts a color from HSV to RGB.
HSV values are in the range [0, 1]. The RGB output is in the range [0, 255]
and are integers.
Parameters
----------
h : float
Hue component, in the range [0, 1].
s : float
Saturation component, in the range [0, 1].
v : float
Value/Brightness component, in the range [0, 1].
Returns
-------
Any
A NumPy array of integers representing the color in RGB format [R, G, B],
with values in the range [0, 255].
"""
color = np.array(colorsys.hsv_to_rgb(h, s, v)) * 255
return color.astype(int)
[docs]
def create_rgb_spectrum(num_steps: int) -> list[int]:
"""
Creates an RGB color spectrum by varying the Hue in HSV space.
The spectrum goes from a hue close to red (0.75) to red (0),
keeping saturation and value at maximum.
Parameters
----------
num_steps : int
The number of steps (colors) to generate in the spectrum.
Returns
-------
list[int]
A list of NumPy arrays, where each array represents an RGB color
([R, G, B]) with integer values in the range [0, 255].
"""
# Values for the HSV components (Hue, Saturation, Value)
h_values = np.linspace(0.75, 0, num_steps)
s = 1.0 # Saturation
v = 1.0 # Value
# Convert HSV to RGB for each hue value
rgb_spectrum = [hsv_to_rgb(h, s, v).astype(np.int32) for h in h_values]
return rgb_spectrum
[docs]
def false_color_scale(channel: Any, color_list: Any, ranges_list: Any) -> Any:
"""
Applies a false color scale to a single-channel image.
Assigns colors from a `color_list` to the pixels of `channel`
based on the ranges specified in `ranges_list`.
Parameters
----------
channel : Any
The single-channel image to which the color scale will be applied.
Expected to be a 2D NumPy array.
color_list : Any
A list of RGB colors ([R, G, B]) to assign to the ranges.
ranges_list : Any
A list of tuples or lists, where each tuple represents a range `(min_val, max_val]`.
Pixel values within this range will be mapped to the corresponding color.
Returns
-------
Any
A new image with the false color scale applied, of integer type and 3 channels.
"""
h, w = channel.shape
new_chroma = np.zeros((h, w, 3))
for i in range(h):
for j in range(w):
for k, r in enumerate(ranges_list):
if r[0] is not None:
if r[0] < channel[i, j] <= r[1]:
new_chroma[i, j] = color_list[k]
new_chroma = new_chroma.astype(np.int32)
return new_chroma
[docs]
def mix_images(img1: Any, img2: Any) -> Any:
"""
Blends two images: an original image and a false-color image.
Pixels in the overlay image that are not completely black (i.e.,
do not have an RGB component sum equal to zero) replace the
corresponding pixels in the original image. If a pixel in `img2` is
black, the original image's pixel is retained.
Parameters
----------
img1 : Any
The original image (e.g., RGB). A NumPy array with 3 channels is expected.
img2 : Any
The image to overlay. A NumPy array with 3 channels is expected.
Returns
-------
Any
The blended image, where non-black pixels from `img2` have replaced
pixels from `img1`.
"""
h, w, _ = img1.shape
new_img = np.zeros(img1.shape, dtype=np.float64)
for i in range(h):
for j in range(w):
if sum(img2[i, j]) != 0:
new_img[i, j] = img2[i, j]
else:
new_img[i, j] = img1[i, j]
return new_img
[docs]
def sort_classifier_results(image, labels, centers, ordered, n_ref, verbose=False):
"""
Sorts clustering results (e.g., K-means) based on
the number of pixels or closeness to LAB reference values.
Parameters
----------
image : npt.NDArray
The original image (flattened if used for clustering) in LAB color space.
labels : npt.NDArray
The cluster identifiers assigned to each pixel.
centers : npt.NDArray
The coordinates (centroids) of the clusters in LAB color space.
ordered : str
The sorting method to use.
Must be "by_pixel" (by number of pixels in each cluster, descending)
or "reference_values" (by closeness to predefined LAB reference values).
n_ref : int
Index of the reference points to use if `ordered` is "reference_values".
Must be 0 or 1, selecting between two sets of LAB references.
verbose : bool, optional
If True, prints detailed information about the sorting process
and the average colors of the clusters. Defaults to False.
Returns
-------
tuple[npt.NDArray, list[npt.NDArray]]
- centers : npt.NDArray
The sorted cluster centroids.
- ordered_images : list[npt.NDArray]
A list of images, where each element is the segmented image
corresponding to a cluster, sorted according to the specified method.
Raises
------
ValueError
If the `ordered` sorting method is invalid.
"""
segmented_data = centers[labels]
segmented_image = segmented_data.reshape(image.shape)
images = []
pixel_num = []
for k in range(centers.shape[0]):
result_image = np.zeros_like(image)
for f in range(image.shape[0]):
if np.allclose(segmented_image[f], centers[k]):
result_image[f] = image[f]
pixel_num.append(np.count_nonzero(result_image))
images.append(result_image)
if ordered == "by_pixel":
# Original behavior
indices_ordenados = np.argsort(pixel_num)[::-1]
centers = centers[indices_ordenados]
ordered_images = sorted(images, key=lambda li: np.count_nonzero(li), reverse=True)
elif ordered == "reference_values":
ref_points = [
np.array([[82.50, 13.58, -3.27], [92.24, 14.87, -5.85], [82.50, 13.58, -3.27]]),
np.array([[30.69, 28.91, -48.47], [60.29, 18.98, -14.80], [81.24, 14.85, -2.15]]),
]
mean_colors = np.array([np.mean(img[img.any(axis=1)], axis=0) for img in images])
if verbose:
print("\n--- Average LAB colors of each cluster ---")
for idx, color in enumerate(mean_colors):
print(f"Cluster {idx}: L={color[0]:.2f}, a={color[1]:.2f}, b={color[2]:.2f}")
print("---------------------------------------------\n")
ordered_indexes = []
available_indexes = list(range(len(mean_colors)))
for i, ref in enumerate(ref_points[n_ref]):
distances = [np.linalg.norm(mean_colors[j] - ref) for j in available_indexes]
closest_idx_in_available = np.argmin(distances)
assigned_cluster = available_indexes[closest_idx_in_available]
if verbose:
print(
f"Reference {i + 1}: Assigned Cluster {assigned_cluster}"
f" (Distance: {distances[closest_idx_in_available]:.4f})"
)
ordered_indexes.append(assigned_cluster)
available_indexes.pop(closest_idx_in_available)
ordered_indexes.append(available_indexes[0])
# TODO: Review reference
indices_ordenados = [ordered_indexes[i] for i in [2, 3, 1, 0]] if n_ref == 1 else ordered_indexes
# The last cluster
if verbose:
print(f"Last cluster assigned: Cluster {available_indexes[0]}\n")
print(indices_ordenados)
centers = centers[indices_ordenados]
ordered_images = [images[i] for i in indices_ordenados]
else:
raise ValueError(f"Invalid order method: {ordered}. Use 'by_pixel' or 'reference_values'.")
return centers, ordered_images
[docs]
def classifier_model(clf_model: str, params: tuple, order: str = "by_pixel", n_ref: int = 1) -> tuple[Any, Any]:
"""
Applies a classification model (K-means, GaussianMixture, or AgglomerativeClustering)
to images and sorts the results.
Parameters
----------
clf_model : {"Kmeans", "GaussianMixture", "AgglomerativeClustering"}
The type of classification model to use.
params : tuple
Specific parameters for the classification model.
- For "Kmeans": (img, k, stop_criteria, number_of_attempts, centroid_initialization_strategy)
- For "GaussianMixture": (img, k)
- For "AgglomerativeClustering": (img, k, linkage)
`img` can be a single image or a list of images.
order : {"by_pixel", "reference_values"}, optional
Method to sort the resulting clusters.
- "by_pixel": sorts by the number of pixels in each cluster (descending).
- "reference_values": sorts by the closeness of centroids to reference LAB values.
Defaults to "by_pixel".
n_ref : int, optional
Reference number (0 or 1) to select the set of reference LAB values
when `order` is "reference_values". Defaults to 1.
Returns
-------
tuple of list
A tuple with:
- centers : list of ndarray
The sorted cluster centroids (or means) for each image.
- ordered_images : list of list of ndarray
A list of lists, where each sublist contains the segmented images
sorted by cluster for the corresponding image.
Raises
------
ValueError
If the `clf_model` classifier type is invalid.
"""
images = params[0]
images = [images] if not isinstance(images, list) else images
lbl_cen = []
if clf_model == "Kmeans":
k, stop_criteria, number_of_attempts, centroid_initialization_strategy = params[1:]
for img in images:
_, labels, centers = cv2.kmeans(
img,
k,
None,
stop_criteria,
number_of_attempts,
centroid_initialization_strategy,
)
lbl_cen.append((labels, centers))
elif clf_model == "GaussianMixture":
k = params[1]
for img in images:
gm = GaussianMixture(n_components=k, init_params="kmeans", random_state=0).fit(img)
labels = gm.predict(img)
centers = gm.means_
lbl_cen.append((labels, centers))
elif clf_model == "AgglomerativeClustering":
k, linkage = params[1], params[-1]
for img in images:
agglom = AgglomerativeClustering(n_clusters=k, linkage=linkage).fit(img)
labels = agglom.labels_
# Compute centers as the mean of the points in each cluster
centers = []
for cluster_label in range(k):
cluster_points = img[labels == cluster_label]
center = np.mean(cluster_points, axis=0)
centers.append(center)
lbl_cen.append((labels, np.array(centers)))
else:
print("Invalid classifier type")
return [list()], [[list(), list()]]
scr = [sort_classifier_results(img, lc[0], lc[1], order, n_ref) for img, lc in zip(images, lbl_cen)]
centers, ordered_images = zip(*scr)
return centers, ordered_images
[docs]
def plot_rgb_3d(images: Any, colors: Any, title: Any) -> None:
"""
Generates a 3D scatter plot of RGB pixels.
Allows visualizing the distribution of colors in 3D space (R, G, B).
Parameters
----------
images : Any | list[Any]
An image or a list of images. Images should be NumPy arrays
representing pixel data (e.g., RGB). If it's a 3D image
(height, width, 3), it will be flattened to (N, 3).
colors : Any | list[Any]
The color or a list of colors for the points in the scatter plot.
Must match the number of images if it's a list.
# Example: colors = [np.array([255, 0, 0]), np.array([0, 255, 0]), np.array([0, 0, 255])]
title : Any | list[Any]
The title or a list of titles for the plot(s).
Must match the number of images if it's a list.
"""
images = images if isinstance(images, list) else [images]
colors = colors if isinstance(colors, list) else [colors]
title = title if isinstance(title, list) else [title]
fig, ax = plt.subplots(1, len(images), figsize=(5 * len(images), 5), subplot_kw={"projection": "3d"})
ax = np.array([ax]) if not isinstance(ax, np.ndarray) else ax
for i, na in enumerate(zip(images, colors, ax)):
img, color, ax = na
if len(img.shape) > 2: # if it's a tensor
img = img.reshape(-1, 3) # flatten to a 2d array
arr = img.T
try:
ax.scatter(arr[1], arr[2], arr[0], c=color, label=f"# px g0: {img.shape[0]}") # RGB 0-1
except ValueError:
# Consider more specific error handling
breakpoint()
ax.set_xlabel("A")
ax.set_ylabel("B")
ax.set_zlabel("L")
ax.legend()
ax.set_title(title[i])
plt.show()
[docs]
def blur_img(image: npt.NDArray) -> npt.NDArray[np.float64]:
"""
Applies a Gaussian filter to an image.
Uses `skimage.filters.gaussian` with a `sigma` of 1 and `truncate` of 1,
resulting in a 3x3 Gaussian kernel.
Parameters
----------
image : npt.NDArray
The input image. Pixel values are expected to be
in the range [0.0, 1.0] (float) or [0, 255] (uint8).
The `skimage.filters.gaussian` function automatically handles data types.
Returns
-------
npt.NDArray[np.float64]
The image with the Gaussian filter applied. The output will be of type float64.
If the input was RGB, the output will also be RGB.
"""
return filters.gaussian(image, channel_axis=-1, truncate=1)
[docs]
def clahe_img(image: npt.NDArray, illumination_axis: int, top_val: float) -> npt.NDArray:
"""
Applies Contrast Limited Adaptive Histogram Equalization (CLAHE)
to the image, specifically to the illumination channel.
The illumination channel is scaled to [0, 255] before applying CLAHE, and then
rescaled back to its original range [0, `top_val`].
Parameters
----------
image : npt.NDArray
The input image. Expected to be an image in a color space
where one channel represents illumination (e.g., the L channel of LAB).
illumination_axis : int
The index of the channel representing illumination in the image (e.g., 0 for L in LAB).
top_val : float
The maximum value the illumination channel can take in its original range
(e.g., 100 for L in LAB).
Returns
-------
npt.NDArray
The image with CLAHE applied to the illumination channel, maintaining float64 data type.
"""
l_ax = image[:, :, illumination_axis]
l_ax *= 255 / top_val
l_ax = l_ax.astype(np.uint8)
clahe = cv2.createCLAHE(clipLimit=5, tileGridSize=(10, 10))
l_ax_eq = clahe.apply(l_ax) # CLAHE range is [0,255] (cv2 range), from lab/skimage.lab L range [0,100]
l_ax_eq = l_ax_eq * (top_val / 255)
image_eq = image.copy()
image_eq[:, :, illumination_axis] = l_ax_eq # float64
return image_eq
[docs]
def pca_img(image: npt.NDArray, comp_num: int) -> npt.NDArray[np.float64]:
"""
Performs Principal Component Analysis (PCA) on the pixels of an image.
Transforms each pixel (interpreted as a point in a 3D space, e.g., RGB or LAB)
to its 3 principal components.
Parameters
----------
image : npt.NDArray
The input image. A NumPy array with pixels expected
in the last dimension (e.g., (height, width, channels)).
comp_num : int
The number of principal components to retain.
Returns
-------
npt.NDArray[np.float64]
A NumPy array representing the transformed image in PCA space,
where each pixel now has 3 values corresponding to its principal components.
The shape will be (N, 3) where N is the total number of pixels.
"""
from sklearn.decomposition import ( # This import is crucial for the function and test
PCA,
)
pca = PCA(n_components=comp_num)
image_pca = pca.fit_transform(image)
return image_pca
[docs]
def find_intersections(x1: np.ndarray, y1: np.ndarray,
x2: np.ndarray, y2: np.ndarray,
distance_threshold: float = 5.0) -> list[tuple[float, float]]:
"""
Find intersection points between two curves, ignoring nearby points.
Parameters
----------
x1 : np.ndarray
X-coordinates of the first curve.
y1 : np.ndarray
Y-coordinates of the first curve.
x2 : np.ndarray
X-coordinates of the second curve.
y2 : np.ndarray
Y-coordinates of the second curve.
distance_threshold : float, optional
Minimum distance between intersections to consider them distinct
(default is 5.0).
Returns
-------
list of tuple of (float, float)
List of (x, y) tuples with the intersections found.
"""
# Create a common finer grid for greater precision
x_min = max(x1.min(), x2.min())
x_max = min(x1.max(), x2.max())
if x_min >= x_max:
return []
x_common = np.linspace(x_min, x_max, 10000)
# Compute the difference between the curves
diff = np.interp(x_common, x1, y1, left=0, right=0) - np.interp(x_common, x2, y2, left=0, right=0)
# Find where the difference changes sign (intersections)
intersections = []
for i in range(len(diff) - 1):
if diff[i] * diff[i + 1] <= 0: # Sign change
# Linear interpolation to find the exact point
if diff[i + 1] - diff[i] != 0:
x_int = x_common[i] - diff[i] * (x_common[i + 1] - x_common[i]) / (diff[i + 1] - diff[i])
else:
x_int = x_common[i]
# Filter points too close to existing intersections
add_point = True
for existing_x, _ in intersections:
if abs(x_int - existing_x) < distance_threshold:
add_point = False
break
if add_point:
intersections.append((x_int, float(np.interp(x_int, x1, y1, left=0, right=0))))
return intersections
[docs]
def filter_intersections_for_console(intersections: list[tuple[float, float]],
distance_threshold: float = 1.0) -> list[tuple[float, float]]:
"""
Filter intersections for console output, removing nearby points.
Parameters
----------
intersections : list of tuple of (float, float)
List of (x, y) intersections.
distance_threshold : float, optional
Minimum distance between points to keep them separate (default is 1.0).
Returns
-------
list of tuple of (float, float)
Filtered list of intersections.
"""
if not intersections:
return []
# Sort by X value
sorted_intersections = sorted(intersections, key=lambda x: x[0])
filtered_intersections = [sorted_intersections[0]]
for i in range(1, len(sorted_intersections)):
x_current, y_current = sorted_intersections[i]
x_previous, y_previous = filtered_intersections[-1]
# Check distance with the last added intersection
if abs(x_current - x_previous) >= distance_threshold:
filtered_intersections.append((x_current, y_current))
return filtered_intersections
[docs]
def detect_number_of_clusters(labels_dir: str) -> int:
"""
Automatically detect the number of clusters used.
Parameters
----------
labels_dir : str
Directory containing label files.
Returns
-------
int
Number of clusters detected.
"""
if not os.path.exists(labels_dir):
warnings.warn(f"Labels directory does not exist: {labels_dir}. Using 3 clusters by default.")
return 3
label_files = [f for f in os.listdir(labels_dir) if f.endswith('_labels.npy')]
if not label_files:
warnings.warn(f"No label files found in {labels_dir}. Using 3 clusters by default.")
return 3
# Load the first label file to detect the number of clusters
first_file = os.path.join(labels_dir, label_files[0])
labels = np.load(first_file)
unique_clusters = np.unique(labels)
# Exclude cluster 0 if it exists
valid_clusters = [c for c in unique_clusters if c != 0]
return len(valid_clusters)
[docs]
def combine_red_blood_cell_clusters(data: dict, n_clusters: int) -> tuple[dict, int]:
"""
Combine clusters 3 and 4 into a single cluster for red blood cells.
Parameters
----------
data : dict
Dictionary with cluster data.
n_clusters : int
Original number of clusters.
Returns
-------
tuple of (dict, int)
Modified data and new number of clusters.
"""
if n_clusters >= 4:
# Combine data from clusters 3 and 4
for component in ['L', 'a', 'b']:
if 4 in data:
data[3][component].extend(data[4][component])
# Remove cluster 4 if it exists
if 4 in data:
del data[4]
return data, n_clusters - 1
return data, n_clusters
[docs]
def get_automatic_names_and_colors(n_clusters: int) -> tuple[dict, dict]:
"""
Automatically generate names and colors based on the number of clusters.
Parameters
----------
n_clusters : int
Number of clusters.
Returns
-------
tuple of (dict, dict)
A tuple containing a names dictionary and a colors dictionary.
"""
# Predefined names for common clusters
predefined_names = {
1: ["Cluster 1"],
2: ["Chromatin", "Cytoplasm"],
3: ["Chromatin", "Cytoplasm", "Red blood cell"],
4: ["Chromatin", "Cytoplasm", "Red blood cell", "Background"],
5: ["Chromatin", "Cytoplasm", "Red blood cell", "Background", "Other"]
}
# Predefined colors (distinctive color scale)
predefined_colors = ['red', 'blue', 'green', 'orange', 'purple',
'brown', 'pink', 'gray', 'olive', 'cyan']
# Get names
if n_clusters in predefined_names:
names = {i + 1: predefined_names[n_clusters][i] for i in range(n_clusters)}
else:
names = {i + 1: f"Cluster {i + 1}" for i in range(n_clusters)}
# Get colors
colors = {i + 1: predefined_colors[i % len(predefined_colors)] for i in range(n_clusters)}
return names, colors
[docs]
def process_image_with_sampling(image_path: str, label_path: str, data: dict,
n_clusters: int, use_sampling: bool = False,
sampling_fraction: float = 0.25) -> None:
"""
Process an image with optional random pixel sampling.
Parameters
----------
image_path : str
Path to the .npy image file.
label_path : str
Path to the .npy label file.
data : dict
Dictionary for storing data.
n_clusters : int
Number of clusters.
use_sampling : bool, optional
If True, use random sampling (default is False).
sampling_fraction : float, optional
Fraction of pixels to sample (default is 0.25).
"""
# Load data
labels = np.load(label_path)
lab_image = np.load(image_path)
h, w = labels.shape
total_pixels = h * w
if use_sampling:
# Create random sampling mask over ALL pixels
final_mask = np.random.choice(
[True, False],
size=total_pixels,
p=[sampling_fraction, 1 - sampling_fraction]
)
# Flatten arrays with sampling
L_flat = lab_image[:, :, 0].flatten()[final_mask]
a_flat = lab_image[:, :, 1].flatten()[final_mask]
b_flat = lab_image[:, :, 2].flatten()[final_mask]
labels_flat = labels.flatten()[final_mask]
else:
# Full processing (no sampling)
L_flat = lab_image[:, :, 0].flatten()
a_flat = lab_image[:, :, 1].flatten()
b_flat = lab_image[:, :, 2].flatten()
labels_flat = labels.flatten()
# Extract data for each cluster
for cluster_id in range(1, n_clusters + 1):
mask = (labels_flat == cluster_id)
data[cluster_id]["L"].extend(L_flat[mask])
data[cluster_id]["a"].extend(a_flat[mask])
data[cluster_id]["b"].extend(b_flat[mask])
[docs]
def compute_histograms(data: dict, n_clusters: int, bins: int = 40) -> dict:
"""
Compute smoothed histograms for all clusters and components.
Parameters
----------
data : dict
Dictionary with cluster data.
n_clusters : int
Number of clusters.
bins : int, optional
Number of bins for the histogram (default is 40).
Returns
-------
dict
Dictionary with smoothed curves per component and cluster.
"""
smoothed_curves = {'L': {}, 'a': {}, 'b': {}}
for cluster_id in range(1, n_clusters + 1):
for comp in ['L', 'a', 'b']:
values = np.array(data[cluster_id][comp])
if len(values) > 0:
# Compute histogram
counts, bins_edges = np.histogram(values, bins=bins)
centers = (bins_edges[:-1] + bins_edges[1:]) / 2
try:
# Smooth with spline
if len(centers) >= 4: # Required for cubic spline
spline = make_interp_spline(centers, counts, k=3)
x_smooth = np.linspace(centers.min(), centers.max(), 500)
y_smooth = spline(x_smooth)
smoothed_curves[comp][cluster_id] = (x_smooth, y_smooth)
else:
# Not enough points, use linear interpolation
smoothed_curves[comp][cluster_id] = (centers, counts)
except Exception as e:
# Fallback to original data
smoothed_curves[comp][cluster_id] = (centers, counts)
return smoothed_curves
[docs]
def compute_all_intersections(smoothed_curves: dict, names: dict,
distance_threshold: float = 6.0) -> dict:
"""
Compute all intersections between cluster pairs for each component.
Parameters
----------
smoothed_curves : dict
Dictionary with curves per component and cluster.
names : dict
Dictionary of cluster names.
distance_threshold : float, optional
Threshold for filtering nearby intersections (default is 6.0).
Returns
-------
dict
Dictionary with intersections per component and cluster pair.
"""
intersections_by_component = {'L': {}, 'a': {}, 'b': {}}
for comp in ['L', 'a', 'b']:
present_clusters = list(smoothed_curves[comp].keys())
# Generate all unique pairs
pairs = [(i, j) for i in present_clusters for j in present_clusters if i < j]
for id1, id2 in pairs:
x1, y1 = smoothed_curves[comp][id1]
x2, y2 = smoothed_curves[comp][id2]
intersections = find_intersections(x1, y1, x2, y2, distance_threshold)
if intersections:
pair_name = f"{names[id1]}-{names[id2]}"
intersections_by_component[comp][pair_name] = intersections
return intersections_by_component
[docs]
def plot_distributions_with_intersections(smoothed_curves: dict,
names: dict, colors: dict,
intersections_by_component: dict,
show_plots: bool = True) -> plt.Figure | None:
"""
Plot distributions with marked intersections.
Parameters
----------
smoothed_curves : dict
Dictionary with smoothed curves.
names : dict
Dictionary of cluster names.
colors : dict
Dictionary of colors per cluster.
intersections_by_component : dict
Computed intersections.
show_plots : bool, optional
If True, display the plots (default is True).
Returns
-------
matplotlib.figure.Figure or None
Matplotlib figure if ``show_plots`` is False, otherwise None.
"""
if not show_plots:
return None
fig, axes = plt.subplots(1, 3, figsize=(20, 6))
fig.suptitle('Frequency Distributions by LAB Component', fontsize=16)
components = ['L', 'a', 'b']
titles = ['L Distribution (Luminosity)',
'a Distribution (Green-Red)',
'b Distribution (Blue-Yellow)']
for idx, comp in enumerate(components):
ax = axes[idx]
# Plot all curves
for cluster_id, (x_smooth, y_smooth) in smoothed_curves[comp].items():
ax.plot(x_smooth, y_smooth, color=colors[cluster_id],
label=names[cluster_id], linewidth=2.5, alpha=0.8)
# Mark intersections for this component
for pair_name, intersections in intersections_by_component[comp].items():
for x_int, y_int in intersections:
# Vertical line at the intersection
ax.axvline(x=x_int, color='gray', linestyle='--', alpha=0.7, linewidth=1)
# Point on the X axis
ax.plot(x_int, 0, 'o', markersize=8, markeredgewidth=2,
markerfacecolor='yellow', markeredgecolor='black')
# Label with the value
ax.text(x_int, ax.get_ylim()[1] * 0.05, f'{x_int:.1f}',
fontsize=9, ha='center', va='bottom', rotation=90,
bbox=dict(boxstyle="round,pad=0.2", facecolor="white", alpha=0.9))
ax.set_xlabel(f'{comp} Value')
ax.set_ylabel('Frequency (pixels)')
ax.set_title(titles[idx])
ax.legend()
ax.grid(True, alpha=0.3)
plt.tight_layout()
if show_plots:
plt.show()
return None
else:
return fig
[docs]
def print_intersection_summary(intersections_by_component: dict,
names: dict, n_clusters: int) -> None:
"""
Print a summary of intersections to the console.
Parameters
----------
intersections_by_component : dict
Dictionary with intersections.
names : dict
Dictionary of cluster names.
n_clusters : int
Number of clusters.
"""
print("\n" + "=" * 60)
print(f"INTERSECTION POINTS SUMMARY ({n_clusters} CLUSTERS)")
print("=" * 60)
for comp in ['L', 'a', 'b']:
print(f"\n--- Component {comp} ---")
if comp in intersections_by_component and intersections_by_component[comp]:
for pair_name, intersections in intersections_by_component[comp].items():
print(f"\nIntersections {pair_name}:")
if intersections:
filtered_intersections = filter_intersections_for_console(
intersections, distance_threshold=3.0
)
for i, (x, y) in enumerate(filtered_intersections, 1):
print(f" Point {i}: x = {x:.2f}")
# Filtering statistics
if intersections and filtered_intersections:
print(f" (Filtered: {len(intersections)} -> {len(filtered_intersections)} points)")
else:
print(" No significant intersections")
else:
print(" No intersections for this component")
[docs]
def analyze_lab_distributions(input_dir: str, labels_dir: str,
use_sampling: bool = False,
sampling_fraction: float = 0.25,
show_plots: bool = True,
bins: int = 40) -> dict:
"""
Main function to analyze LAB distributions and compute intersections.
Parameters
----------
input_dir : str
Directory with .npy image files.
labels_dir : str
Directory with .npy label files.
use_sampling : bool, optional
If True, use random sampling (default is False).
sampling_fraction : float, optional
Fraction of pixels to sample (default is 0.25).
show_plots : bool, optional
If True, display plots (default is True).
bins : int, optional
Number of bins for histograms (default is 40).
Returns
-------
dict
Dictionary with all analysis results including keys: ``'data'``,
``'n_clusters'``, ``'names'``, ``'colors'``, ``'smoothed_curves'``,
``'intersections_by_component'``, ``'num_images_processed'``,
``'use_sampling'``, ``'sampling_fraction'``.
"""
# Detect number of clusters
n_clusters = detect_number_of_clusters(labels_dir)
print(f"Number of clusters detected: {n_clusters}")
# Initialize data
data = {cluster_id: {"L": [], "a": [], "b": []} for cluster_id in range(1, n_clusters + 1)}
# Process images
if not os.path.exists(labels_dir):
raise ValueError(f"No label files found in {labels_dir}")
label_files = [f for f in os.listdir(labels_dir) if f.endswith('_labels.npy')]
if not label_files:
raise ValueError(f"No label files found in {labels_dir}")
print(f"Processing {len(label_files)} images...")
for i, file in enumerate(label_files, 1):
base_name = file.replace("_labels.npy", "")
label_path = os.path.join(labels_dir, file)
image_path = os.path.join(input_dir, f"{base_name}.npy")
if not os.path.exists(image_path):
warnings.warn(f"Image not found: {image_path}")
continue
process_image_with_sampling(image_path, label_path, data, n_clusters,
use_sampling, sampling_fraction)
if i % 10 == 0:
print(f" Processed {i}/{len(label_files)} images")
# Combine clusters if necessary
data, n_clusters = combine_red_blood_cell_clusters(data, n_clusters)
# Get names and colors
names, colors = get_automatic_names_and_colors(n_clusters)
# Compute smoothed histograms
print("Computing smoothed histograms...")
smoothed_curves = compute_histograms(data, n_clusters, bins)
# Compute intersections
print("Computing intersections...")
intersections_by_component = compute_all_intersections(
smoothed_curves, names
)
# Plot (if requested)
if show_plots:
print("Generating plots...")
plot_distributions_with_intersections(
smoothed_curves, names, colors, intersections_by_component, True
)
# Print summary
print_intersection_summary(intersections_by_component, names, n_clusters)
# Return all results
return {
'data': data,
'n_clusters': n_clusters,
'names': names,
'colors': colors,
'smoothed_curves': smoothed_curves,
'intersections_by_component': intersections_by_component,
'num_images_processed': len(label_files),
'use_sampling': use_sampling,
'sampling_fraction': sampling_fraction if use_sampling else 1.0
}
[docs]
def process_npy_images(input_folder: str, output_folder: str,
kernel_size: int = 3, process_rgb: bool = True) -> None:
"""
Process all .npy images in a folder by applying a median filter.
Parameters
----------
input_folder : str
Path to the folder with .npy files.
output_folder : str
Path where processed images will be saved.
kernel_size : int, optional
Kernel size (must be odd) (default is 3).
process_rgb : bool, optional
If True, process as RGB (3 channels); if False, as grayscale
(default is True).
Raises
------
ValueError
If folders do not exist or kernel_size is invalid.
"""
if not os.path.exists(input_folder):
raise ValueError(f"Input folder does not exist: {input_folder}")
if kernel_size <= 0 or kernel_size % 2 == 0:
raise ValueError("kernel_size must be a positive odd number")
# Create output folder if it does not exist
os.makedirs(output_folder, exist_ok=True)
# Get .npy files
files = [f for f in os.listdir(input_folder) if f.lower().endswith('.npy')]
for filename in files:
try:
input_path = os.path.join(input_folder, filename)
image = np.load(input_path)
if process_rgb:
# Verify it is RGB
if len(image.shape) != 3 or image.shape[2] != 3:
print(f"Warning: {filename} is not RGB. Skipping...")
continue
filtered_image = apply_median_filter_rgb(image, kernel_size)
else:
# Process as grayscale
filtered_image = apply_median_filter(image, kernel_size)
# Save result
base_name = os.path.splitext(filename)[0]
output_path = os.path.join(output_folder, f"{base_name}_filtered.npy")
np.save(output_path, filtered_image)
except Exception as e:
print(f"Error processing {filename}: {str(e)}")
[docs]
def create_test_matrix(size: int = 5, kind: str = 'ordered') -> np.ndarray:
"""
Create test matrices to verify the median filter.
Parameters
----------
size : int, optional
Size of the square matrix (default is 5).
kind : str, optional
Type of matrix: ``'ordered'``, ``'random'``, ``'stepped'``, or
``'edges'`` (default is ``'ordered'``).
Returns
-------
np.ndarray
Test matrix.
Raises
------
ValueError
If the matrix type is not valid.
"""
if kind == 'ordered':
# Matrix with sequentially ordered values
return np.array([[i + j * size for i in range(size)]
for j in range(size)], dtype=np.float32)
elif kind == 'random':
# Matrix with random values
np.random.seed(42) # For reproducibility
return np.random.randint(0, 100, (size, size)).astype(np.float32)
elif kind == 'stepped':
# Matrix with stepped values (good for verifying filters)
matrix = np.zeros((size, size), dtype=np.float32)
for i in range(size):
for j in range(size):
matrix[i, j] = max(i, j)
return matrix
elif kind == 'edges':
# Matrix with high edges and low center (for testing edge preservation)
matrix = np.ones((size, size), dtype=np.float32) * 10
border = size // 4
matrix[:border, :] = 100 # Top edge
matrix[-border:, :] = 100 # Bottom edge
matrix[:, :border] = 100 # Left edge
matrix[:, -border:] = 100 # Right edge
return matrix
else:
raise ValueError(f"Invalid matrix type: {kind}")
"""
Functions for Noise Overlap Index (NOI) computation and histogram accumulation.
"""
import numpy as np
import os
[docs]
def accumulate_histograms(input_dir: str, labels_dir: str, n_clusters: int,
component: str = 'L', bins: int = 40,
use_sampling: bool = False,
sampling_fraction: float = 0.25) -> tuple[dict, tuple]:
"""
Accumulate histograms per cluster for the chosen component ('L', 'a', 'b').
Parameters
----------
input_dir : str
Folder with .npy images in LAB format.
labels_dir : str
Folder with K-means label files.
n_clusters : int
Number of detected clusters.
component : str, optional
Component to use: ``'L'``, ``'a'``, or ``'b'`` (default is ``'L'``).
bins : int, optional
Number of bins for the histogram (default is 40).
use_sampling : bool, optional
If True, use random sampling to speed up processing (default is False).
sampling_fraction : float, optional
Fraction of pixels to sample if ``use_sampling`` is True
(default is 0.25).
Returns
-------
tuple of (dict, tuple)
A tuple ``(accumulators, range)`` where ``accumulators`` is a dict
with histograms per cluster and ``range`` is a tuple
``(range_min, range_max)`` of the range used.
"""
# Prepare accumulators
accumulators = {}
# Fixed range for Lab
if component == 'L':
range_min, range_max = 0.0, 100.0
else: # 'a' or 'b'
range_min, range_max = -128.0, 127.0
comp_idx = {'L': 0, 'a': 1, 'b': 2}[component]
# Get label files
label_files = [f for f in os.listdir(labels_dir)
if f.endswith('_labels.npy')]
for file in label_files:
base = file.replace("_labels.npy", "")
label_path = os.path.join(labels_dir, file)
image_path = os.path.join(input_dir, f"{base}.npy")
if not os.path.exists(image_path):
continue
labels = np.load(label_path)
lab = np.load(image_path)
# Extract values from the selected component
vals = lab[:, :, comp_idx].flatten()
labs_flat = labels.flatten()
# Apply sampling if enabled
if use_sampling:
total = vals.size
mask = np.random.choice([True, False], size=total,
p=[sampling_fraction, 1 - sampling_fraction])
vals = vals[mask]
labs_flat = labs_flat[mask]
# Accumulate histograms per cluster
for cid in range(n_clusters):
maskc = (labs_flat == cid)
if np.any(maskc):
counts, _ = np.histogram(vals[maskc], bins=bins,
range=(range_min, range_max))
if cid not in accumulators:
accumulators[cid] = np.zeros(bins, dtype=np.float64)
accumulators[cid] += counts
# Combine clusters 3 and 4 if they exist (for the specific malaria case)
if n_clusters >= 5 and 3 in accumulators and 4 in accumulators:
accumulators[3] = accumulators[3] + accumulators[4]
del accumulators[4]
return accumulators, (range_min, range_max)
[docs]
def compute_noi(hist1: np.ndarray, hist2: np.ndarray,
range_values: tuple, bins: int = 40) -> float:
"""
Compute the NOI (overlap area) between two accumulated histograms.
Parameters
----------
hist1 : np.ndarray
Histogram of the first cluster.
hist2 : np.ndarray
Histogram of the second cluster.
range_values : tuple of (float, float)
Tuple ``(range_min, range_max)`` of the histogram range.
bins : int, optional
Number of histogram bins (default is 40).
Returns
-------
float
NOI value between 0 and 1.
"""
range_min, range_max = range_values
bin_width = (range_max - range_min) / bins
# Normalize to probability densities
area1 = hist1.sum()
area2 = hist2.sum()
if area1 == 0 or area2 == 0:
return 0.0
p1 = hist1 / (area1 * bin_width)
p2 = hist2 / (area2 * bin_width)
# Compute overlap area
overlap = np.sum(np.minimum(p1, p2)) * bin_width
return float(overlap)
[docs]
def simulate_distributions(mu1: float, sigma1: float, mu2: float, sigma2: float,
bins: int = 40, n_samples: int = 10000) -> tuple:
"""
Simulate two normal distributions and compute the NOI.
Useful for unit tests.
Parameters
----------
mu1 : float
Mean of the first distribution.
sigma1 : float
Standard deviation of the first distribution.
mu2 : float
Mean of the second distribution.
sigma2 : float
Standard deviation of the second distribution.
bins : int, optional
Number of bins (default is 40).
n_samples : int, optional
Number of samples to generate (default is 10000).
Returns
-------
tuple of (np.ndarray, np.ndarray, tuple, float)
A tuple ``(hist1, hist2, range, computed_noi)``.
"""
range_min, range_max = 0.0, 100.0 # For L component
# Generate samples
samples1 = np.clip(np.random.normal(mu1, sigma1, n_samples),
range_min, range_max)
samples2 = np.clip(np.random.normal(mu2, sigma2, n_samples),
range_min, range_max)
# Create histograms
hist1, _ = np.histogram(samples1, bins=bins, range=(range_min, range_max))
hist2, _ = np.histogram(samples2, bins=bins, range=(range_min, range_max))
# Compute NOI
noi = compute_noi(hist1, hist2, (range_min, range_max), bins=bins)
return hist1, hist2, (range_min, range_max), noi
[docs]
def analytical_noi_gaussians(mu1: float, sigma1: float, mu2: float, sigma2: float,
range_min: float = 0.0, range_max: float = 100.0,
n_points: int = 10000) -> float:
"""
Analytically compute the NOI for two normal distributions.
Parameters
----------
mu1 : float
Mean of the first distribution.
sigma1 : float
Standard deviation of the first distribution.
mu2 : float
Mean of the second distribution.
sigma2 : float
Standard deviation of the second distribution.
range_min : float, optional
Minimum value of the range (default is 0.0).
range_max : float, optional
Maximum value of the range (default is 100.0).
n_points : int, optional
Number of points for numerical integration (default is 10000).
Returns
-------
float
Analytically computed NOI.
"""
# Create grid for integration
x_start = max(mu1 - 4 * sigma1, mu2 - 4 * sigma2, range_min)
x_end = min(mu1 + 4 * sigma1, mu2 + 4 * sigma2, range_max)
# No overlap between the distributions' relevant ranges
if x_start >= x_end:
return 0.0
x = np.linspace(x_start, x_end, n_points)
# Normal PDFs
pdf1 = (1 / (sigma1 * np.sqrt(2 * np.pi))) * np.exp(-0.5 * ((x - mu1) / sigma1) ** 2)
pdf2 = (1 / (sigma2 * np.sqrt(2 * np.pi))) * np.exp(-0.5 * ((x - mu2) / sigma2) ** 2)
# Normalize within the range
mask = (x >= range_min) & (x <= range_max)
if not np.any(mask):
return 0.0
pdf1 = pdf1[mask]
pdf2 = pdf2[mask]
x = x[mask]
# Normalize to area = 1 within the range
pdf1 = pdf1 / np.trapezoid(pdf1, x)
pdf2 = pdf2 / np.trapezoid(pdf2, x)
# Compute NOI
analytical_noi = np.trapezoid(np.minimum(pdf1, pdf2), x)
return float(analytical_noi)