【将图片格式转换为调色盘模式——语义分割】

在做语义分割时，复现GitHub代码，出现了语义类标签随机的问题，最终定位到是转换为调试盘时没有统一调色盘，为了节省时间，将代码写为多进程，现将完整代码分享。

转换为调色盘格式

from PIL import Image, ImageDraw, ImageFont
import os
import cv2
import numpy as np
import multiprocessing

def get_all_colorsjr(infile):#获取主要颜色
    colors_set = set()
    names = os.listdir(infile)
    for name in names:
        if not (name.endswith('jpg') or name.endswith('png')):
            print('errname',name)
            continue
        lab_name = os.path.join(infile,name)
        image = Image.open(lab_name)
        # print(image.mode)
        # small_image = image.resize((10, 10))
        small_image = image

        width = small_image.size[0]
        height = small_image.size[1]
        for i in range(0, width):
            for j in range(0, height):
                a = small_image.getpixel((i, j))
                colors_set.add(a)
        colors_set_list = list(colors_set)
        colors_set_list.sort()
        colors_dict = {}
        for colors in colors_set_list:
            colors_dict[colors] = colors_set_list.index(colors)

        print('colors_dict',colors_dict)
    return colors_dict


def OsDistance(vector1, vector2):  # 计算欧式距离
    n1 = float(vector1[0])
    n2 = float(vector1[1])
    n3 = float(vector1[2])
    m1 = float(vector2[0])
    m2 = float(vector2[1])
    m3 = float(vector2[2])

    distance = ((n1 - m1) ** 2 + (n2 - m2) ** 2 + (n3 - m3) ** 2) ** 0.5
    return distance

def calculate_distance( a,c): #找到最相近的颜色，并取出标签
    m=[]
    for ci in c:
        d=OsDistance(a,ci)
        m.append(d)
    t=m.index(min(m))
    return t



def RGB2P(img_name,palette,colors):
    img = Image.open(img_name)

    # cv_img = cv2.imread(img_name)

    if img.mode !='P':
        print('61img_name',img_name,img.mode)
        # img = Image.open(img_name)
        width = img.size[0]
        height = img.size[1]
        temp = 1
        # img = img.convert("P", palette=Image.ADAPTIVE)
        imgP = img.convert("P")

        imgP.putpalette(palette)
        for i in range(0, width):
            for j in range(0, height):
                a=img.getpixel((i,j))

                try:
                    b=calculate_distance(a,colors)
                    # b = colors.index(a)
                    imgP.putpixel((i, j), colors[b])
                except Exception as e:
                    print('errimg_name',img_name,e)
                    raise a

        # imgP = img.convert("P")
        # img = img.convert("P",palette=Image.ADAPTIVE)
        # img.palette.colors = colors_dict
        # img.putpalette(palette)
        # imgP.save(img_name[:-4]+'_PP.png')
        imgP.save(img_name)
        # print('ok')
        # imgPP = Image.open(img_name[:-4]+'_PP.png')
        # print(imgP.palette.colors)
    else:
        print('81', img_name,img.mode)
        # raise img_name


if __name__ == '__main__':





    colors = [(0, 0, 0), (85, 255, 0), (53, 119, 181), (255, 0, 255), (245, 128, 6), (255, 255, 127), (255, 0, 0),(255, 255, 255)]

    # colors_dict = {(0, 0, 0):0,(85, 255, 0):1,(53, 119, 181):2,(255, 0, 255 ):3,(245, 128, 6 ):4,(255, 255, 127):5,(255, 0, 0 ):6,(255, 255, 255):7}

    palette = []
    for i in range(8):
        palette.extend((i, i, i))
    # 原来的，图片颜色有问题，绿色和红色反的
    palette[:3 * 8] = np.array([(0, 0, 0), (85, 255, 0), (53, 119, 181), (255, 0, 255), (245, 128, 6), (255, 255, 127), (255, 0, 0),(255, 255, 255)], dtype='uint8').flatten()

    # # 翻转之后,列表标签超出，无法编码，不能用
    # palette[:3 * 8] = np.array(
    #     [(0, 0, 0), (0, 255, 85), (181, 119, 53,), (255, 0, 255), (6, 128, 245), (127, 255, 255), (0, 0, 255),(255, 255, 255)], dtype='uint8').flatten()

    path_my = r''

    # get_all_colorsjr(path_my)

    pool = multiprocessing.Pool(processes=20)
    for img_name in os.listdir(path_my):
        img_name = os.path.join(path_my, img_name)
        # print('img_name',img_name)
        pool.apply_async(func=RGB2P, args=(img_name,palette,colors))

    pool.close()
    pool.join()  # 调用join之前，先调用close函数，否则会出错。执行完close后不会有新的进程加入到pool,join函数等待所有子进程结束
    print('gg')

检查调色盘是否正确（检查语义分割的类标签是否正确）

通过下面代码打印的set_P 判断生成的调色盘有没有问题，即判断生成的类有几种，每个类分别对应什么颜色。

from PIL import Image
import os,cv2
import multiprocessing

dict = {}
set_P = set()
count = 0


def Check(img_name):
    global dict
    global set_P
    global count
    print('img_name', img_name)
    img = Image.open(img_name)
    cv_img = cv2.imread(img_name)

    # # print(cv_img.shape[2])
    # if cv_img.shape[2]!=3:
    #     print(cv_img.shape)

    if img.mode !='P':
        print('23img.mode !=P',img_name,img.mode)
        # continue
        # # img = img.convert('P')
        # img = img.convert("P", palette=Image.ADAPTIVE)
        # # img.save(img_name)
        # print(img.mode)
    elif img.mode =='P':
        # print(img.size)
        print('31img.palette.colors',img.palette.colors)
        for x in  range(img.size[0]):
            for y in range(img.size[1]):
                if img.getpixel((x,y)) !=1 or True:
                    # print(img.getpixel((x,y)))
                    # print(cv_img[y][x])
                    dict[str(cv_img[y][x])] = img.getpixel((x,y))
                    set_P.add(str(img.getpixel((x,y)))+"______"+str(cv_img[y][x]))

        print('dict',count,len(dict),dict)
        print('set_P',count,len(set_P),set_P)
    count+=1

        # break

from mv_similar import simJr

if __name__ == '__main__':

    path_my = r''
    # path_my = path_ori


    for img_name in os.listdir(path_my):
        img_name = os.path.join(path_my, img_name)

    pool = multiprocessing.Pool(processes=30)
    for img_name in os.listdir(path_my):
        img_name = os.path.join(path_my, img_name)
        # print('img_name',img_name)
        pool.apply_async(func=Check, args=(img_name,))
        # pool.apply_async(func=simJr, args=('img_name','s'))

    pool.close()
    pool.join()  # 调用join之前，先调用close函数，否则会出错。执行完close后不会有新的进程加入到pool,join函数等待所有子进程结束