python用直方图规定化实现图像风格转换

以下内容需要直方图均衡化、规定化知识

均衡化：https://blog.csdn.net/macunshi/article/details/79815870

规定化：https://blog.csdn.net/macunshi/article/details/79819263

直方图均衡化应用：

图像直方图均衡化能拉伸灰度图，让像素值均匀分布在0,255之间，使图像看起来不会太亮或太暗，常用于图像增强；

直方图规定化应用：

举个例子，当我们需要对多张图像进行拼接时，我们希望这些图片的亮度、饱和度保持一致，事实上就是让它们的直方图分布一致，这时就需要直方图规定化。

直方图规定化与均衡化的思想一致，事实上就是找到各个灰度级别的映射关系。具体实现的过程中一般会选一个参考图像记为A，找到A的直方图与目标图像的直方图的映射关系，从而找到目标图像的像素以A为“参考”时的映射关系。

具体实现可参考文中链接（看完茅塞顿开）

基于python利用直方图规定化统一图像风格

参考图像

原始图像（第一行）/处理后的图像（第二行）

源码：

import os
import cv2
import numpy as np
 
def get_map(Hist):
    # 计算概率分布Pr
    sum_Hist = sum(Hist)
    Pr = Hist/sum_Hist
    # 计算累计概率Sk
    Sk = []
    temp_sum = 0
    for n in Pr:
        temp_sum = temp_sum + n
        Sk.append(temp_sum)
    Sk = np.array(Sk)
    # 计算映射关系img_map
    img_map = []
    for m in range(256):
        temp_map = int(255*Sk[m] + 0.5)
        img_map.append(temp_map)
    img_map = np.array(img_map)
    return img_map
 
def get_off_map(map_): # 计算反向映射，寻找最小期望
    map_2 = list(map_)
    off_map = []
    temp_pre = 0 # 如果循环开始就找不到映射时，默认映射为0
    for n in range(256):
        try:
            temp1 = map_2.index(n)
            temp_pre = temp1
        except BaseException:
            temp1 = temp_pre # 找不到映射关系时，近似取向前最近的有效映射值
        off_map.append(temp1)
    off_map = np.array(off_map)
    return off_map
 
def get_infer_map(infer_img):
    infer_Hist_b = cv2.calcHist([infer_img], [0], None, [256], [0,255])
    infer_Hist_g = cv2.calcHist([infer_img], [1], None, [256], [0,255])
    infer_Hist_r = cv2.calcHist([infer_img], [2], None, [256], [0,255])
    infer_b_map = get_map(infer_Hist_b)
    infer_g_map = get_map(infer_Hist_g)
    infer_r_map = get_map(infer_Hist_r)
    infer_b_off_map = get_off_map(infer_b_map)
    infer_g_off_map = get_off_map(infer_g_map)
    infer_r_off_map = get_off_map(infer_r_map)
    return [infer_b_off_map, infer_g_off_map, infer_r_off_map]
 
def get_finalmap(org_map, infer_off_map): # 计算原始图像到最终输出图像的映射关系
    org_map = list(org_map)
    infer_off_map = list(infer_off_map)
    final_map = []
    for n in range(256):
        temp1 = org_map[n]
        temp2 = infer_off_map[temp1]
        final_map.append(temp2)
    final_map = np.array(final_map)
    return final_map
 
def get_newimg(img_org, org2infer_maps):
    w, h, _ = img_org.shape
    b, g ,r =cv2.split(img_org)
    for i in range(w):
        for j in range(h):
            temp1 = b[i,j]
            b[i,j] = org2infer_maps[0][temp1]
    for i in range(w):
        for j in range(h):
            temp1 = g[i,j]
            g[i,j] = org2infer_maps[1][temp1]
    for i in range(w):
        for j in range(h):
            temp1 = r[i,j]
            r[i,j] = org2infer_maps[2][temp1]
    newimg = cv2.merge([b,g,r])
    return newimg
 
def get_new_img(img_org, infer_map):
    org_Hist_b = cv2.calcHist([img_org], [0], None, [256], [0,255])
    org_Hist_g = cv2.calcHist([img_org], [1], None, [256], [0,255])
    org_Hist_r = cv2.calcHist([img_org], [2], None, [256], [0,255])
    org_b_map = get_map(org_Hist_b)
    org_g_map = get_map(org_Hist_g)
    org_r_map = get_map(org_Hist_r)
    org2infer_map_b = get_finalmap(org_b_map, infer_map[0])
    org2infer_map_g = get_finalmap(org_g_map, infer_map[1])
    org2infer_map_r = get_finalmap(org_r_map, infer_map[2])
    return get_newimg(img_org, [org2infer_map_b, org2infer_map_g, org2infer_map_r])
 
if __name__ == "__main__":
    dstroot = './imgs'
    infer_img_path = './abc.png'
    infer_img = cv2.imread(infer_img_path)
    outroot = './out1'
    infer_map = get_infer_map(infer_img) # 计算参考映射关系
    dstlist = os.listdir(dstroot)
    for n in dstlist:
        img_path = os.path.join(dstroot, n)
        print(img_path)
        img_org = cv2.imread(img_path)
        new_img = get_new_img(img_org, infer_map) # 根据映射关系获得新的图像
        new_path = os.path.join(outroot, n)
        cv2.imwrite(new_path, new_img)