sora_python/app/service/outfit_matcher/foco.py

import operator
from numba import jit
import numpy as np
from skimage import transform, color


@jit(nopython=True)
def h_unit(h):
    """
    The unit value (1-15) of hue
    """
    for i, u in [(0, 1), (14, 2), (33, 3), (44, 4), (62, 5), (88, 6), (165, 7), (194, 8), (219, 9), (241, 10),
                 (266, 11), (292, 12), (307, 13), (327, 14), (344, 15), (360, 1)]:
        if h <= i:
            return u
    return 0


@jit(nopython=True)
def s_unit(s):
    """
    The unit value (1-8) of saturation
    """
    for i, u in [(0, 1), (5, 1), (10, 2), (14, 3), (34, 4), (53, 5), (63, 6), (80, 7), (100, 8)]:
        if s <= i:
            return u
    return 0


@jit(nopython=True)
def b_unit(b):
    """
    The unit value (1-6) of brightness
    """
    for i, u in [(0, 1), (20, 1), (39, 2), (55, 3), (72, 4), (90, 5), (100, 6)]:
        if b <= i:
            return u
    return 0


@jit(nopython=True)
def encode_foco(hsb):
    h, s, b = hsb
    theta = (h - 1) / 15 * np.pi * 2
    return np.array([np.cos(theta), np.sin(theta), (s - 1) / 7, (b - 1) / 5])


@jit(nopython=True)
def is_white(hsb_unit):
    """
    Whether an hsb color unit is white
    """
    _, s, b = hsb_unit
    return s == 1 and b == 6


@jit(nopython=True)
def is_rice(hsb_unit):
    """
    Whether an hsb color unit is rice
    """
    h, s, b = hsb_unit
    return h in [3, 4, 5] and s in [2, 3, 4] and b in [5, 6]


@jit(nopython=True)
def is_blue_netrual(hsb_unit):
    """
    Whether an hsb color is blue netrual
    """
    h, s, b = hsb_unit
    return h in [9, 10] and s in [2, 3, 4, 5, 6] and b in [2, 3, 4]


@jit(nopython=True)
def is_pastel(hsb_unit):
    """
    Whether an hsb color unit is pastel
    """
    _, s, b = hsb_unit
    return s == 1 and b in [4, 5]


@jit(nopython=True)
def is_black(hsb_unit):
    """
    Whether an hsb color unit is black
    """
    _, _, b = hsb_unit
    return b == 1


@jit(nopython=True)
def is_gray(hsb_unit):
    """
    Whether an hsb color unit is gray
    """
    _, s, b = hsb_unit
    return s == 1 and b in [2, 3]


@jit(nopython=True)
def foco_merge(hsb_foco):
    """
    Merge multiple black/white/gray foco unit to one
    """
    if is_white(hsb_foco): return (1, 1, 6)
    if is_black(hsb_foco): return (1, 1, 1)
    if is_gray(hsb_foco): return (1, hsb_foco[1], hsb_foco[2])
    return hsb_foco[0], hsb_foco[1], hsb_foco[2]


def color_foco(img):
    """
    transform an hsb image to color units

    input: (height, width, 3) array
    output: an uint8 array with same shape
    """
    img = img.copy()
    h, w, c = img.shape
    # print(h, w, c)
    assert c == 3
    for i in range(h):
        for j in range(w):
            img[i, j, 0] = h_unit(img[i, j, 0])
            img[i, j, 1] = s_unit(img[i, j, 1])
            img[i, j, 2] = b_unit(img[i, j, 2])
    return img.astype(np.uint8)


@jit(nopython=True)
def is_ignore(hsb, eps=1):
    """
    Ignore white background
    """
    _, s, b = hsb
    return s < eps and b > 100 - eps


@jit(nopython=True)
def ignore_list(hsb_img, ignore=is_ignore):
    """
    Given an hsb image (ranges 360, 100, 100), output a dict of ignored pixel locations

    Input:
    is_ignore: a lambda that determines whether a hsb pixel should be ignored
    Output:
    set of ignore pixel coordinate
    """
    h, w, _ = hsb_img.shape
    ign = set()
    for i in range(h):
        for j in range(w):
            if ignore(hsb_img[i, j, :]):
                ign.add((i, j))
    return ign


def color_histogram(img, merge=None, threshold=None, ignore=None):
    """
    color histogram of an image
    """
    h, w, c = img.shape
    # print(h, w, c)
    if merge is None:
        merge = foco_merge
    if threshold is None:
        threshold = h + w
    if ignore is None:
        ignore = {}
    hist = {}
    assert c == 3
    for i in range(h):
        for j in range(w):
            if (i, j) in ignore:
                continue
            k = merge(tuple(img[i, j]))
            if k in hist:
                hist[k] += 1
            else:
                hist[k] = 1

    return {k: v for k, v in hist.items() if v > threshold}


def main_colors(img, n=1, frequency=False, hist=False, merge=True):
    """
    return the list of main colors of a hsb image.

    img: a hsb image
    n: number of main colors to return
    frequency: whether return freqency
    """
    if hist:
        hist = img
    else:
        hist = color_histogram(color_foco(img), ignore=ignore_list(img))
    if merge:
        newhist = {}
        oldk = {}
        ks = sorted(hist.items(), key=operator.itemgetter(1), reverse=True)
        # print(ks)
        for k, v in ks:
            if k in oldk:
                continue
            near = {kk: hist[kk] for kk in hist if np.abs(np.sum(np.abs(np.array(k) - np.array(kk)))) <= 2}
            # print(near)
            newhist[k] = sum(v for k, v in near.items())
            for kk in near:
                oldk[kk] = None
                hist.pop(kk)
        hist = newhist
    items = sorted(hist.items(), key=operator.itemgetter(1), reverse=True)[:n]
    return items if frequency else [k for k, v in items]


def rgb2hsb(rgb_img):
    """
    Transform an rgb image (unit 8 np array, ranges 255, 255, 255) to hsb (float np array, ranges 360, 100, 100)
    """
    rgb_img = np.array(rgb_img).astype('uint8')
    return color.rgb2hsv(rgb_img) * np.array([360, 100, 100]).reshape(1, 1, 3)


def extract_main_colors(img, n=5):
    """
    Args:
        img: Numpy array (height, width, channel)
        n: number of main colors wants to extract

    return:
        Features of main colors: Numpy array (n, 5)
    """
    # Convert to hsb
    img = img.astype("uint8")
    height = img.shape[0]
    width = img.shape[1]
    ratio = (512 * 512) / (height * width)
    if ratio < 1.0:
        img = transform.resize(img, (int(height * ratio), int(width * ratio)))
    img = color.rgb2hsv(img) * np.array([360, 100, 100]).reshape(1, 1, 3)

    # Extract main colors
    cf = main_colors(img, n, frequency=True)
    s = sum(f for c, f in cf)
    features = np.zeros((n, 5))
    for i, (c, f) in enumerate(cf):
        features[i, :4] = encode_foco(c)
        features[i, 4] = f / s
    return features