语义分割中的类别不平衡的权重计算

这是5幅图，加上背景共5类。 

 

可以参考这篇文章https://blog.csdn.net/u012426298/article/details/81232386

对于一个多类别图片数据库，每个类别都会有一个class frequency, 该类别像素数目除以数据库总像素数目, 求出所有class frequency 的median 值，除以该类别对应的frequency 得到weight：

weight=median(weights)/weights

#coding:utf-8
from __future__ import print_function
import os
import numpy as np
import cv2
w,h=512,512
def find_pic(img,array_list,n_class,pixs):
    img_sum = np.sum(img == array_list, axis=-1)
    pix_numbers=img_sum.reshape(-1).tolist().count(3)
    if pix_numbers:
        pixs+=pix_numbers
        n_class+=1
    return pixs,n_class
def compute_class(pixs,n_class):
    return pixs/(n_class*w*h)
def frequence():
    # images_path = './trainannot_visual'
    images_path = './visual_example'
    red = np.array([0, 0, 128])
    yellow = np.array([0, 128, 128])
    green = np.array([0, 128, 0])
    blue = np.array([128, 0, 0])
    back_gro = np.array([0, 0, 0])

    images_list_path = [os.path.join(images_path,i) for i in os.listdir(images_path)]

    n_red=0
    red_pixs=0

    n_yellow = 0
    yellow_pixs = 0

    n_green= 0
    green_pixs = 0

    n_blue = 0
    blue_pixs = 0

    n_back = 0
    back_pixs = 0

    for count,image_path in enumerate(images_list_path):
        print('{}image'.format(count))

        img=cv2.imread(image_path)

        red_pixs, n_red=find_pic(img,red,n_red,red_pixs)

        yellow_pixs,n_yellow = find_pic(img, yellow,n_yellow,yellow_pixs)

        green_pixs,n_green = find_pic(img, green,n_green,green_pixs)
        blue_pixs,n_blue = find_pic(img, blue,n_blue,blue_pixs)
        #
        back_pixs,n_back = find_pic(img, back_gro,n_back,back_pixs)

    print(red_pixs, n_red)
    print(yellow_pixs,n_yellow)
    print(green_pixs, n_green)
    print(blue_pixs, n_blue)
    print(back_pixs, n_back)

    f_class_red=compute_class(red_pixs,n_red)
    f_class_yellow = compute_class(yellow_pixs, n_yellow)
    f_class_green = compute_class(green_pixs, n_green)
    f_class_blue = compute_class(blue_pixs, n_blue)
    f_class_back = compute_class(back_pixs, n_back)
    print(f_class_red,f_class_yellow,f_class_green,f_class_blue,f_class_back)

    f_class=[f_class_red,f_class_yellow,f_class_green,f_class_blue,f_class_back]

    f_class_median=np.median(np.array(f_class))
    print(f_class_median)
    print(f_class_median/np.array(f_class))
if __name__ == '__main__':
    frequence()

这样可以保证占比小的class, 权重大于1， 占比大的class， 权重小于1， 达到balancing的效果.