OpenCV 中的向量思维

OpenCV 中得矩阵和向量思维
比如下面去除背景算法，源码用c++写的，我用python重写了一遍，
参考1
用opencv 从实现如下，核心代码如下

def reduce_background(img):
    win_size = 50
    
    nh = int(img.shape[0]/win_size)+1
    nw = int(img.shape[1]/win_size)+1

    for i in range(nh):
        for j in range(nw):
            # print (i,j)
            rstart = j*win_size
            cstart = i*win_size

            if rstart+win_size>img.shape[0]-1:
                rend = img.shape[0]-1
            else:
                rend = rstart+win_size

            if cstart+win_size>img.shape[1]-1:
                cend = img.shape[1]-1
            else:
                cend = cstart+win_size

            summ = float(win_size*win_size)

            # sub = img[rstart:rend,cstart:cend]
            # cv2.imshow("original"+str(i)+str(j),sub)

            ar = 0
            ag = 0
            ab = 0

            for r in range(rstart,rend):
                for c in range(cstart,cend):
                    ar += img[r,c][0]/summ
                    ag += img[r,c][1]/summ
                    ab += img[r,c][2]/summ

            # print(ar,ag,ab)
            agray = 0.3*ar+0.59*ag+0.11*ab
            # print(agray)

            if agray<0.1:
                bright = 1.0
            else:
                bright = 250.0/agray

            contr = 1.3

            for r in range(rstart,rend):
                for c in range(cstart,cend):
                    
                    vv = (ar +(img[r,c][0]-ar)*contr)*bright
                    img[r,c][0] = clamp(vv)

                    vv = (ag +(img[r,c][1]-ag)*contr)*bright
                    img[r,c][1] = clamp(vv)

                    vv = (ab +(img[r,c][2]-ab)*contr)*bright
                    img[r,c][2] = clamp(vv)

发现大量的时间用在for循环里，参考了网上的建议，把代码改写如下

def reduce_background(img):
    """
    reduce background of any image by (size*size) square method 
    version 2:use numpy MAT accelate much faster than python approch

    """
    win_size = 50
    print (img.shape)
    
    nh = int(img.shape[0]/win_size)+1
    nw = int(img.shape[1]/win_size)+1

    for i in range(nw):
        for j in range(nh):
            # print (i,j)
            rstart = j*win_size
            cstart = i*win_size

            if rstart + win_size>img.shape[0]-1:
                rend = img.shape[0]-1
            else:
                rend = rstart + win_size

            if cstart + win_size>img.shape[1]-1:
                cend = img.shape[1]-1
            else:
                cend = cstart + win_size

            summ = float(win_size*win_size)
            sub = img[rstart:rend,cstart:cend]

            ar = 0 # average color of red
            ag = 0 # average color of green
            ab = 0 # average color of blue

            rr = sub[:,:,0]
            gg = sub[:,:,1]
            bb = sub[:,:,2]

            ar = cv2.mean(rr)
            ag = cv2.mean(gg)
            ab = cv2.mean(bb)

            agray = 0.3*ar[0]+0.59*ag[0]+0.11*ab[0] #average color of grey

            maxv = np.max(sub)
            minv = np.min(sub)

            s = 1 - minv/maxv
            v = maxv/agray

            if v>0.3:
                bright = 2
            else:
                bright = 250.0/agray

            contr = 1.3

            # enhanse image
            rr = np.clip((ar[0]+(rr-ar[0])*contr)*bright,0,255)
            gg = np.clip((ag[0]+(gg-ag[0])*contr)*bright,0,255)
            bb = np.clip((ab[0]+(bb-ab[0])*contr)*bright,0,255)

            # update src image
            img[rstart:rend,cstart:cend][:,:,0]=rr
            img[rstart:rend,cstart:cend][:,:,1]=gg
            img[rstart:rend,cstart:cend][:,:,2]=bb

    return img

矩阵计算

用numpy矩阵函数执行效率大幅提升
for— 用向量运算 numpy ±*（）等直接向量矩阵运算

rr = np.clip((ar[0]+(rr-ar[0])*contr)*bright,0,255)

子矩阵

sub = img[rstart:rend,cstart:cend] #子矩阵 或子图像
rr = sub[:,:,0] #”取出图像的红色分量
img[rstart:rend,cstart:cend][:,:,0]=rr  # 填回数据
# 

最大最小值
直接用np.max() 或 np.min() 我一开始用错成 minMaxLoc 函数走了很多弯路
当然还有 np.clip()等函数 不一一详述

总结
用numpy中的内置函数效率能提高，维护更方便