Python 颜色迁移

watersink 2022-05-23 我要评论

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)
    """