app.py

import numpy as np
import gradio as gr
from PIL import Image
import cv2
from skimage import color
from sklearn.cluster import KMeans
from typing import Tuple


def proc(img: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
    assert img.shape[-1] == 3
    k_size = 11
    sigma = 11
    blurred = cv2.GaussianBlur(img, (k_size, k_size), sigma)
    blurred_small = cv2.resize(blurred, (80, 80))
    labs = color.rgb2lab(blurred_small)
    lab_vectors = labs.reshape(-1, 3)

    num_colors = 5
    num_bins = 10
    km = KMeans(n_clusters=num_colors)
    km.fit(lab_vectors)
    centroid_labs = km.cluster_centers_  # N x (L, a, b)
    centroid_labs = np.array(
        sorted(centroid_labs, key=lambda x: x[1] ** 2 + x[2] ** 2)
    )  # sort by L
    seeds = np.log(np.arange(0, 100, num_bins) + num_bins)
    ls = seeds * 100 / seeds[-1]
    centroid_ls = np.clip(ls, 0, 100).reshape(1, num_bins, 1).repeat(num_colors, axis=0)
    centroid_abs = centroid_labs[:, np.newaxis, 1:].repeat(num_bins, axis=1)
    centroid_labs = np.concatenate([centroid_ls, centroid_abs], axis=-1).reshape(
        num_colors, num_bins, 3
    )

    unique_indices = [0] + [
        i
        for i in range(1, num_colors)
        if np.linalg.norm(centroid_labs[i] - centroid_labs[i - 1]) > 10
    ]
    centroid_labs = centroid_labs[unique_indices, :, :]

    centroid_rgbs = (color.lab2rgb(centroid_labs) * 255).astype(np.uint8)
    centroid_rgb_vis = cv2.resize(
        centroid_rgbs,
        (int(img.shape[0] / num_colors * num_bins), img.shape[0]),
        interpolation=cv2.INTER_NEAREST,
    )
    return centroid_rgb_vis


demo = gr.Interface(fn=proc, inputs="image", outputs="image")

if __name__ == "__main__":
    demo.launch()