Python 颜色迁移
watersink 人气:0前言
reinhard算法:Color Transfer between Images,作者Erik Reinhard
welsh算法:Transferring Color to Greyscale Images,作者Tomihisa Welsh
应用场景
人像图换肤色,风景图颜色迁移
出发点
- RGB三通道有很强的关联性,而做颜色的改变同时恰当地改变三通道比较困难。
- 需要寻找三通道互不相关的也就是正交的颜色空间,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,从而能够较好的保持原图的自然效果。三个通道分别代表:亮度,黄蓝通道,红绿通道。
reinhard算法流程
- 输入变换图,颜色参考图,将其都从bgr空间转化为lab空间
- 分别计算变换图,参考图在lab空间的均值,方差
- (变换图lab - 变换图均值)/变换图方差 *参考图方差 + 参考图均值
- 变换图lab空间转化为bgr空间,输出结果
welsh算法流程
- 输入变换图,颜色参考图,将其都从bgr空间转化为lab空间
- 定义随机参考点个数segment,领域空间大小window_size,加权系数ratio。从参考图片中随机选择segment个样本点,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,求这2个的加权W,W = L* ratio+ σ*(1-ratio)。这样就可以得到segment个W,以及与其一一对应的a通道,b通道对应位置的数值。
- 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,保证后续的像素匹配正确进行
- 对变换图进行逐像素扫描,对每个像素,计算其权值W,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,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。
- 将变换图从Lab空间转化到bgr空间。
Reinhard VS welsh
- Reinhard 操作简单,高效,速度快很多。
- welsh算法涉及到了参考图的W的计算,如果是参考图固定且已知的场景,这一步可以放入初始化中。如果不是这样的场景,那么这一步的计算也是很费时的。
- welsh整体速度慢很多,主要由于求方差造成。
- welsh的输出效果,受随机参考点个数以及位置的影响,每次的结果都会有差异。
- welsh的效果会有种涂抹不均匀的感觉,Reinhard 则没有这种问题。
代码实现
Reinhard
def color_trans_reinhard(in_img, ref_img, in_mask_lists=[None], ref_mask_lists=[None]): ref_img_lab = cv2.cvtColor(ref_img, cv2.COLOR_BGR2LAB) in_img_lab = cv2.cvtColor(in_img, cv2.COLOR_BGR2LAB) in_avg = np.ones(in_img.shape, np.float32) in_std = np.ones(in_img.shape, np.float32) ref_avg = np.ones(in_img.shape, np.float32) ref_std = np.ones(in_img.shape, np.float32) mask_all = 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 = cv2.meanStdDev(in_img_lab, mask=in_mask) np.copyto(in_avg, in_avg_tmp.reshape(1,1,-1), where=np.expand_dims(in_mask,2)!=0) #numpy.copyto(destination, source) np.copyto(in_std, in_std_tmp.reshape(1,1,-1), where=np.expand_dims(in_mask,2)!=0) ref_avg_tmp, ref_std_tmp = cv2.meanStdDev(ref_img_lab, mask=ref_mask) np.copyto(ref_avg, ref_avg_tmp.reshape(1,1,-1), where=np.expand_dims(in_mask,2)!=0) #numpy.copyto(destination, source) np.copyto(ref_std, ref_std_tmp.reshape(1,1,-1), where=np.expand_dims(in_mask,2)!=0) #mask mask_all[in_mask!=0] = 1 in_std[in_std==0] =1 #避免除数为0的情况 transfered_lab = (in_img_lab - in_avg)/(in_std) *ref_std + ref_avg transfered_lab[transfered_lab<0] = 0 transfered_lab[transfered_lab>255] = 255 out_img = 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=mask_all==0) return out_img """ #img1 = cv2.imread("imgs/1.png") #img2 = cv2.imread("imgs/2.png") #img1 = cv2.imread("welsh22/1.png", 1) #img2 = cv2.imread("welsh22/2.png", 1) img1 = cv2.imread("welsh22/gray.jpg", 1) img2 = 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 = cv2.imread("ab.jpeg") img2 = cv2.imread("hsy.jpeg") mask1 = cv2.imread("ab_parsing.jpg", 0) mask1[mask1<128]=0 mask1[mask1>=128]=255 mask2 = cv2.imread("hsy_parsing.jpg", 0) mask2[mask2<128]=0 mask2[mask2>=128]=255 cv2.imwrite("out.jpg", color_trans_reinhard(img1, img2, [mask1], [mask2]))
Welsh代码
改进点
- 主要是去掉for循环操作。
- 将计算一个领域内的std,使用均值滤波+numpy实现近似替换。差别目测看不出。
- 修改参考图的weight,全部int化,只保留不一样的weight,实际测试大概150个左右的weight就可以。
- 修改最近weight查找思路,使用numpy减法操作+argmin,替换2分查找。
- 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 = max(pixel[1] - window_size, 0) window_right = min(pixel[1] + window_size + 1, width) window_top = max(pixel[0] - window_size, 0) window_bottom = min(pixel[0] + window_size + 1, height) window_slice = 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=[None]): weight_list = [] pixel_a_list = [] pixel_b_list = [] ref_img_mask = 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 = np.bitwise_or(x, ref_img_mask) ref_img_l_mean = cv2.blur(ref_img_l, (window_size, window_size)) ref_img_l_std = np.sqrt(cv2.blur(np.power((ref_img_l - ref_img_l_mean), 2), (window_size, window_size))) for _ in range(segment): height_index = np.random.randint(ref_img_height) width_index = np.random.randint(ref_img_width) pixel = [height_index, width_index] #[x,y] if ref_img_mask[pixel[0], pixel[1]] == 0: continue pixel_light = ref_img_l[pixel[0], pixel[1]] pixel_a = ref_img_a[pixel[0], pixel[1]] pixel_b = ref_img_b[pixel[0], pixel[1]] #pixel_std = get_domain_std(ref_img_l, pixel, ref_img_height, ref_img_width, window_size) pixel_std = ref_img_l_std[height_index, width_index] weight_value = int(pixel_light * ratio + 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=[None], ref_mask_lists=[None]): start = time.time() #参考图 ref_img_height, ref_img_width, ref_img_channel = ref_img.shape window_size=5 #窗口大小 segment= 10000#随机点个数 ratio=0.5 #求weight的比例系数 ref_img_lab = cv2.cvtColor(ref_img, cv2.COLOR_BGR2Lab) ref_img_l, ref_img_a, ref_img_b = cv2.split(ref_img_lab) #计算参考图weight ref_img_weight_array, ref_img_pixel_a_array, ref_img_pixel_b_array = 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 = np.max(ref_img_l), np.min(ref_img_l) #输入图 in_img_height, in_img_width, in_img_channel = in_img.shape in_img_lab = cv2.cvtColor(in_img, cv2.COLOR_BGR2LAB) # 获取灰度图像的亮度信息; in_img_l, in_img_a, in_img_b = cv2.split(in_img_lab) in_img_max_pixel, in_img_min_pixel = np.max(in_img_l), np.min(in_img_l) pixel_ratio = (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; in_img_l = ref_img_min_pixel + (in_img_l - in_img_min_pixel) * pixel_ratio in_img_l = in_img_l.astype(np.uint8) in_img_l_mean = cv2.blur(in_img_l, (window_size, window_size)) in_img_l_std = np.sqrt(cv2.blur(np.power((in_img_l - in_img_l_mean), 2), (window_size, window_size))) in_img_weight_pixel = ratio * in_img_l + (1 - ratio) * in_img_l_std nearest_pixel_index = np.argmin(np.abs(ref_img_weight_array.reshape(1,1,-1) - np.expand_dims(in_img_weight_pixel, 2)), axis=2).astype(np.float32) in_img_a = cv2.remap(ref_img_pixel_a_array.reshape(1, -1), nearest_pixel_index, np.zeros_like(nearest_pixel_index, np.float32), interpolation=cv2.INTER_LINEAR) in_img_b = cv2.remap(ref_img_pixel_b_array.reshape(1, -1), nearest_pixel_index, np.zeros_like(nearest_pixel_index, np.float32), interpolation=cv2.INTER_LINEAR) merge_img = cv2.merge([in_img_l, in_img_a, in_img_b]) bgr_img = cv2.cvtColor(merge_img, cv2.COLOR_LAB2BGR) mask_all = 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 = np.bitwise_or(x, mask_all) mask_all = cv2.merge([mask_all, mask_all, mask_all]) np.copyto(bgr_img, in_img, where=mask_all==0) end = time.time() print("time", end-start) return bgr_img if __name__ == '__main__': # 创建参考图像的分析类; #ref_img = cv2.imread("consult.jpg") #ref_img = cv2.imread("2.png") ref_img = cv2.imread("../imgs/2.png") # 读取灰度图像;opencv默认读取的是3通道的,不需要我们扩展通道; #in_img = cv2.imread("gray.jpg") #in_img = cv2.imread("1.png") in_img = cv2.imread("../imgs/1.png") bgr_img = color_trans_welsh(in_img, ref_img) cv2.imwrite("out_ren.jpg", bgr_img) """ ref_img = cv2.imread("../hsy.jpeg") ref_mask = cv2.imread("../hsy_parsing.jpg", 0) ref_mask[ref_mask<128] = 0 ref_mask[ref_mask>=128] = 255 in_img = cv2.imread("../ab.jpeg") in_mask = cv2.imread("../ab_parsing.jpg", 0) in_mask[in_mask<128] = 0 in_mask[in_mask>=128] = 255 bgr_img = color_trans_welsh(in_img, ref_img, in_mask_lists=[in_mask], ref_mask_lists=[ref_mask]) cv2.imwrite("bgr.jpg", bgr_img) """
效果对比
从左到右,分别为原图,参考图,reinhard效果,welsh效果
从左到右,分别为原图,原图皮肤mask,参考图,参考图皮肤mask,reinhard效果,welsh效果
加载全部内容