不要谈什么天分&＃xff0c;运气&＃xff0c;

你需要的是一个截稿日&＃xff0c;

以及一个不交稿就能打爆你狗头的人&＃xff0c;

然后你就会被自己的才华吓到。         -------查理·布洛克

reinhard算法&＃xff1a;Color Transfer between Images&＃xff0c;作者Erik Reinhard

welsh算法&＃xff1a;Transferring Color to Greyscale Images&＃xff0c;作者Tomihisa Welsh

应用场景&＃xff1a;

人像图换肤色&＃xff0c;风景图颜色迁移

出发点&＃xff1a;

1. RGB三通道有很强的关联性&＃xff0c;而做颜色的改变同时恰当地改变三通道比较困难。
2. 需要寻找三通道互不相关的也就是正交的颜色空间&＃xff0c;作者想到了Ruderman等人提出的lαβ颜色空间。三个轴向正交意味着改变任何一个通道都不影响其他通道&＃xff0c;从而能够较好的保持原图的自然效果。三个通道分别代表&＃xff1a;亮度&＃xff0c;黄蓝通道&＃xff0c;红绿通道。

reinhard算法流程&＃xff1a;

1. 输入变换图&＃xff0c;颜色参考图&＃xff0c;将其都从bgr空间转化为lab空间
2. 分别计算变换图&＃xff0c;参考图在lab空间的均值&＃xff0c;方差
3. (变换图lab - 变换图均值)/变换图方差 *参考图方差 &＃43; 参考图均值
4. 变换图lab空间转化为bgr空间&＃xff0c;输出结果

welsh算法流程&＃xff1a;

1. 输入变换图&＃xff0c;颜色参考图&＃xff0c;将其都从bgr空间转化为lab空间
2. 定义随机参考点个数segment&＃xff0c;领域空间大小window_size&＃xff0c;加权系数ratio。从参考图片中随机选择segment个样本点&＃xff0c;将这些样本点的像素亮度值L和L空间window_size领域内得方差σ保存起来&＃xff0c;求这2个的加权W&＃xff0c;W &＃61; L* ratio&＃43; σ*(1-ratio)。这样就可以得到segment个W&＃xff0c;以及与其一一对应的a通道&＃xff0c;b通道对应位置的数值。
3. 对变换图的L通道基于颜色参考图的L通道进行亮度重映射&＃xff0c;保证后续的像素匹配正确进行
4. 对变换图进行逐像素扫描&＃xff0c;对每个像素&＃xff0c;计算其权值W&＃xff0c;计算方式和上面一样。然后在第二步得到的样本点中找到与其权值最接近的参考点&＃xff0c;并将该点的a通道和b通道的值赋给变换图的a通道和b通道。
5. 将变换图从Lab空间转化到bgr空间。

Reinhard VS welsh:

1. Reinhard 操作简单&＃xff0c;高效&＃xff0c;速度快很多。
2. welsh算法涉及到了参考图的W的计算&＃xff0c;如果是参考图固定且已知的场景&＃xff0c;这一步可以放入初始化中。如果不是这样的场景&＃xff0c;那么这一步的计算也是很费时的。
3. welsh整体速度慢很多&＃xff0c;主要由于求方差造成。
4. welsh的输出效果&＃xff0c;受随机参考点个数以及位置的影响&＃xff0c;每次的结果都会有差异。
5. welsh的效果会有种涂抹不均匀的感觉&＃xff0c;Reinhard 则没有这种问题。

代码实现&＃xff1a;

Reinhard:

def color_trans_reinhard(in_img, ref_img, in_mask_lists&＃61;[None], ref_mask_lists&＃61;[None]):ref_img_lab &＃61; cv2.cvtColor(ref_img, cv2.COLOR_BGR2LAB)in_img_lab &＃61; cv2.cvtColor(in_img, cv2.COLOR_BGR2LAB)in_avg &＃61; np.ones(in_img.shape, np.float32)in_std &＃61; np.ones(in_img.shape, np.float32)ref_avg &＃61; np.ones(in_img.shape, np.float32)ref_std &＃61; np.ones(in_img.shape, np.float32)mask_all &＃61; np.zeros(in_img.shape, np.float32)for in_mask, ref_mask in zip(in_mask_lists, ref_mask_lists):#mask,取值为 0, 255, shape[height,width]in_avg_tmp, in_std_tmp &＃61; cv2.meanStdDev(in_img_lab, mask&＃61;in_mask)np.copyto(in_avg, in_avg_tmp.reshape(1,1,-1), where&＃61;np.expand_dims(in_mask,2)!&＃61;0) #numpy.copyto(destination, source)np.copyto(in_std, in_std_tmp.reshape(1,1,-1), where&＃61;np.expand_dims(in_mask,2)!&＃61;0) ref_avg_tmp, ref_std_tmp &＃61; cv2.meanStdDev(ref_img_lab, mask&＃61;ref_mask)np.copyto(ref_avg, ref_avg_tmp.reshape(1,1,-1), where&＃61;np.expand_dims(in_mask,2)!&＃61;0) #numpy.copyto(destination, source)np.copyto(ref_std, ref_std_tmp.reshape(1,1,-1), where&＃61;np.expand_dims(in_mask,2)!&＃61;0) #maskmask_all[in_mask!&＃61;0] &＃61; 1in_std[in_std&＃61;&＃61;0] &＃61;1 #避免除数为0的情况transfered_lab &＃61; (in_img_lab - in_avg)/(in_std) *ref_std &＃43; ref_avg transfered_lab[transfered_lab<0] &＃61; 0transfered_lab[transfered_lab>255] &＃61; 255out_img &＃61; cv2.cvtColor(transfered_lab.astype(np.uint8), cv2.COLOR_LAB2BGR)if in_mask_lists[0] is not None and ref_mask_lists[0] is not None:np.copyto(out_img, in_img, where&＃61;mask_all&＃61;&＃61;0) return out_img"""
#img1 &＃61; cv2.imread("imgs/1.png")
#img2 &＃61; cv2.imread("imgs/2.png")
#img1 &＃61; cv2.imread("welsh22/1.png", 1)
#img2 &＃61; cv2.imread("welsh22/2.png", 1)
img1 &＃61; cv2.imread("welsh22/gray.jpg", 1)
img2 &＃61; cv2.imread("welsh22/consult.jpg", 1)
cv2.imwrite("out.jpg", color_trans_reinhard(img1, img2, [np.ones(img1.shape[:-1],np.uint8)*255], [np.ones(img2.shape[:-1],np.uint8)*255]))"""
img1 &＃61; cv2.imread("ab.jpeg?s=#34;)
img2 &＃61; cv2.imread("hsy.jpeg?s=#34;)
mask1 &＃61; cv2.imread("ab_parsing.jpg", 0)
mask1[mask1<128]&＃61;0
mask1[mask1>&＃61;128]&＃61;255
mask2 &＃61; cv2.imread("hsy_parsing.jpg", 0)
mask2[mask2<128]&＃61;0
mask2[mask2>&＃61;128]&＃61;255
cv2.imwrite("out.jpg", color_trans_reinhard(img1, img2, [mask1], [mask2]))


Welsh代码&＃xff1a;

改进点&＃xff0c;

1. 主要是去掉for循环操作。
2. 将计算一个领域内的std&＃xff0c;使用均值滤波&＃43;numpy实现近似替换。差别目测看不出。
3. 修改参考图的weight&＃xff0c;全部int化&＃xff0c;只保留不一样的weight&＃xff0c;实际测试大概150个左右的weight就可以。
4. 修改最近weight查找思路&＃xff0c;使用numpy减法操作&＃43;argmin&＃xff0c;替换2分查找。
5. 整体速度比原始代码快18倍。

def get_domain_std(img_l, pixel, height, width, window_size):window_left &＃61; max(pixel[1] - window_size, 0)window_right &＃61; min(pixel[1] &＃43; window_size &＃43; 1, width)window_top &＃61; max(pixel[0] - window_size, 0)window_bottom &＃61; min(pixel[0] &＃43; window_size &＃43; 1, height)window_slice &＃61; img_l[window_top: window_bottom, window_left: window_right]return np.std(window_slice)def get_weight_pixel(ref_img_l, ref_img_a, ref_img_b, ref_img_height, ref_img_width, segment, window_size, ratio, ref_mask_lists&＃61;[None]):weight_list &＃61; []pixel_a_list &＃61; []pixel_b_list &＃61; []ref_img_mask &＃61; np.ones((ref_img_height, ref_img_width), np.uint8)if ref_mask_lists[0] is not None:for x in ref_mask_lists:ref_img_mask &＃61; np.bitwise_or(x, ref_img_mask)ref_img_l_mean &＃61; cv2.blur(ref_img_l, (window_size, window_size))ref_img_l_std &＃61; np.sqrt(cv2.blur(np.power((ref_img_l - ref_img_l_mean), 2), (window_size, window_size)))for _ in range(segment):height_index &＃61; np.random.randint(ref_img_height)width_index &＃61; np.random.randint(ref_img_width)pixel &＃61; [height_index, width_index] #[x,y]if ref_img_mask[pixel[0], pixel[1]] &＃61;&＃61; 0:continuepixel_light &＃61; ref_img_l[pixel[0], pixel[1]]pixel_a &＃61; ref_img_a[pixel[0], pixel[1]]pixel_b &＃61; ref_img_b[pixel[0], pixel[1]]#pixel_std &＃61; get_domain_std(ref_img_l, pixel, ref_img_height, ref_img_width, window_size)pixel_std &＃61; ref_img_l_std[height_index, width_index]weight_value &＃61; int(pixel_light * ratio &＃43; pixel_std * (1 - ratio))if weight_value not in weight_list:weight_list.append(weight_value)pixel_a_list.append(pixel_a)pixel_b_list.append(pixel_b) return np.array(weight_list), np.array(pixel_a_list), np.array(pixel_b_list)def color_trans_welsh(in_img, ref_img, in_mask_lists&＃61;[None], ref_mask_lists&＃61;[None]):start &＃61; time.time()#参考图ref_img_height, ref_img_width, ref_img_channel &＃61; ref_img.shapewindow_size&＃61;5 #窗口大小segment&＃61; 10000#随机点个数ratio&＃61;0.5 #求weight的比例系数ref_img_lab &＃61; cv2.cvtColor(ref_img, cv2.COLOR_BGR2Lab)ref_img_l, ref_img_a, ref_img_b &＃61; cv2.split(ref_img_lab)#计算参考图weightref_img_weight_array, ref_img_pixel_a_array, ref_img_pixel_b_array &＃61; get_weight_pixel(ref_img_l, ref_img_a, ref_img_b, ref_img_height, ref_img_width, segment, window_size, ratio, ref_mask_lists)ref_img_max_pixel, ref_img_min_pixel &＃61; np.max(ref_img_l), np.min(ref_img_l)#输入图in_img_height, in_img_width, in_img_channel &＃61; in_img.shapein_img_lab &＃61; cv2.cvtColor(in_img, cv2.COLOR_BGR2LAB)# 获取灰度图像的亮度信息&＃xff1b;in_img_l, in_img_a, in_img_b &＃61; cv2.split(in_img_lab)in_img_max_pixel, in_img_min_pixel &＃61; np.max(in_img_l), np.min(in_img_l)pixel_ratio &＃61; (ref_img_max_pixel - ref_img_min_pixel) / (in_img_max_pixel - in_img_min_pixel)# 把输入图像的亮度值映射到参考图像范围内&＃xff1b;in_img_l &＃61; ref_img_min_pixel &＃43; (in_img_l - in_img_min_pixel) * pixel_ratioin_img_l &＃61; in_img_l.astype(np.uint8)in_img_l_mean &＃61; cv2.blur(in_img_l, (window_size, window_size))in_img_l_std &＃61; np.sqrt(cv2.blur(np.power((in_img_l - in_img_l_mean), 2), (window_size, window_size)))in_img_weight_pixel &＃61; ratio * in_img_l &＃43; (1 - ratio) * in_img_l_stdnearest_pixel_index &＃61; np.argmin(np.abs(ref_img_weight_array.reshape(1,1,-1) - np.expand_dims(in_img_weight_pixel, 2)), axis&＃61;2).astype(np.float32)in_img_a &＃61; cv2.remap(ref_img_pixel_a_array.reshape(1, -1), nearest_pixel_index, np.zeros_like(nearest_pixel_index, np.float32), interpolation&＃61;cv2.INTER_LINEAR)in_img_b &＃61; cv2.remap(ref_img_pixel_b_array.reshape(1, -1), nearest_pixel_index, np.zeros_like(nearest_pixel_index, np.float32), interpolation&＃61;cv2.INTER_LINEAR)merge_img &＃61; cv2.merge([in_img_l, in_img_a, in_img_b])bgr_img &＃61; cv2.cvtColor(merge_img, cv2.COLOR_LAB2BGR)mask_all &＃61; np.zeros(in_img.shape[:-1], np.int32)if in_mask_lists[0] is not None and ref_mask_lists[0] is not None:for x in in_mask_lists:mask_all &＃61; np.bitwise_or(x, mask_all)mask_all &＃61; cv2.merge([mask_all, mask_all, mask_all])np.copyto(bgr_img, in_img, where&＃61;mask_all&＃61;&＃61;0) end &＃61; time.time()print("time", end-start)return bgr_imgif __name__ &＃61;&＃61; &＃39;__main__&＃39;:# 创建参考图像的分析类&＃xff1b;#ref_img &＃61; cv2.imread("consult.jpg")#ref_img &＃61; cv2.imread("2.png")ref_img &＃61; cv2.imread("../imgs/2.png")# 读取灰度图像&＃xff1b;opencv默认读取的是3通道的&＃xff0c;不需要我们扩展通道&＃xff1b;#in_img &＃61; cv2.imread("gray.jpg")#in_img &＃61; cv2.imread("1.png")in_img &＃61; cv2.imread("../imgs/1.png")bgr_img &＃61; color_trans_welsh(in_img, ref_img)cv2.imwrite("out_ren.jpg", bgr_img)"""ref_img &＃61; cv2.imread("../hsy.jpeg?s=#34;)ref_mask &＃61; cv2.imread("../hsy_parsing.jpg", 0)ref_mask[ref_mask<128] &＃61; 0ref_mask[ref_mask>&＃61;128] &＃61; 255in_img &＃61; cv2.imread("../ab.jpeg?s=#34;)in_mask &＃61; cv2.imread("../ab_parsing.jpg", 0)in_mask[in_mask<128] &＃61; 0in_mask[in_mask>&＃61;128] &＃61; 255bgr_img &＃61; color_trans_welsh(in_img, ref_img, in_mask_lists&＃61;[in_mask], ref_mask_lists&＃61;[ref_mask])cv2.imwrite("bgr.jpg", bgr_img)"""

效果对比&＃xff1a;

从左到右&＃xff0c;分别为原图&＃xff0c;参考图&＃xff0c;reinhard效果&＃xff0c;welsh效果 

 从左到右&＃xff0c;分别为原图&＃xff0c;原图皮肤mask&＃xff0c;参考图&＃xff0c;参考图皮肤mask&＃xff0c;reinhard效果&＃xff0c;welsh效果