""" @author: cuny @file: ThinFace.py @time: 2022/7/2 15:50 @description: 瘦脸算法，用到了图像局部平移法 先使用人脸关键点检测，然后再使用图像局部平移法 需要注意的是，这部分不会包含dlib人脸关键点检测，因为考虑到模型载入的问题 """ import cv2 import math import numpy as np class TranslationWarp(object): """ 本类包含瘦脸算法，由于瘦脸算法包含了很多个版本，所以以类的方式呈现 前两个算法没什么好讲的，网上资料很多 第三个采用numpy内部的自定义函数处理，在处理速度上有一些提升 最后采用cv2.map算法，处理速度大幅度提升 """ # 瘦脸 @staticmethod def localTranslationWarp(srcImg, startX, startY, endX, endY, radius): # 双线性插值法 def BilinearInsert(src, ux, uy): w, h, c = src.shape if c == 3: x1 = int(ux) x2 = x1 + 1 y1 = int(uy) y2 = y1 + 1 part1 = src[y1, x1].astype(np.float64) * (float(x2) - ux) * (float(y2) - uy) part2 = src[y1, x2].astype(np.float64) * (ux - float(x1)) * (float(y2) - uy) part3 = src[y2, x1].astype(np.float64) * (float(x2) - ux) * (uy - float(y1)) part4 = src[y2, x2].astype(np.float64) * (ux - float(x1)) * (uy - float(y1)) insertValue = part1 + part2 + part3 + part4 return insertValue.astype(np.int8) ddradius = float(radius * radius) # 圆的半径 copyImg = srcImg.copy() # copy后的图像矩阵 # 计算公式中的|m-c|^2 ddmc = (endX - startX) * (endX - startX) + (endY - startY) * (endY - startY) H, W, C = srcImg.shape # 获取图像的形状 for i in range(W): for j in range(H): # # 计算该点是否在形变圆的范围之内 # # 优化，第一步，直接判断是会在（startX,startY)的矩阵框中 if math.fabs(i - startX) > radius and math.fabs(j - startY) > radius: continue distance = (i - startX) * (i - startX) + (j - startY) * (j - startY) if distance < ddradius: # 计算出（i,j）坐标的原坐标 # 计算公式中右边平方号里的部分 ratio = (ddradius - distance) / (ddradius - distance + ddmc) ratio = ratio * ratio # 映射原位置 UX = i - ratio * (endX - startX) UY = j - ratio * (endY - startY) # 根据双线性插值法得到UX，UY的值 # start_ = time.time() value = BilinearInsert(srcImg, UX, UY) # print(f"双线性插值耗时;{time.time() - start_}") # 改变当前 i ，j的值 copyImg[j, i] = value return copyImg # 瘦脸pro1, 限制了for循环的遍历次数 @staticmethod def localTranslationWarpLimitFor(srcImg, startP: np.matrix, endP: np.matrix, radius: float): startX, startY = startP[0, 0], startP[0, 1] endX, endY = endP[0, 0], endP[0, 1] # 双线性插值法 def BilinearInsert(src, ux, uy): w, h, c = src.shape if c == 3: x1 = int(ux) x2 = x1 + 1 y1 = int(uy) y2 = y1 + 1 part1 = src[y1, x1].astype(np.float64) * (float(x2) - ux) * (float(y2) - uy) part2 = src[y1, x2].astype(np.float64) * (ux - float(x1)) * (float(y2) - uy) part3 = src[y2, x1].astype(np.float64) * (float(x2) - ux) * (uy - float(y1)) part4 = src[y2, x2].astype(np.float64) * (ux - float(x1)) * (uy - float(y1)) insertValue = part1 + part2 + part3 + part4 return insertValue.astype(np.int8) ddradius = float(radius * radius) # 圆的半径 copyImg = srcImg.copy() # copy后的图像矩阵 # 计算公式中的|m-c|^2 ddmc = (endX - startX) ** 2 + (endY - startY) ** 2 # 计算正方形的左上角起始点 startTX, startTY = (startX - math.floor(radius + 1), startY - math.floor((radius + 1))) # 计算正方形的右下角的结束点 endTX, endTY = (startX + math.floor(radius + 1), startY + math.floor((radius + 1))) # 剪切srcImg srcImg = srcImg[startTY: endTY + 1, startTX: endTX + 1, :] # db.cv_show(srcImg) # 裁剪后的图像相当于在x,y都减少了startX - math.floor(radius + 1) # 原本的endX, endY在切后的坐标点 endX, endY = (endX - startX + math.floor(radius + 1), endY - startY + math.floor(radius + 1)) # 原本的startX, startY剪切后的坐标点 startX, startY = (math.floor(radius + 1), math.floor(radius + 1)) H, W, C = srcImg.shape # 获取图像的形状 for i in range(W): for j in range(H): # 计算该点是否在形变圆的范围之内 # 优化，第一步，直接判断是会在（startX,startY)的矩阵框中 # if math.fabs(i - startX) > radius and math.fabs(j - startY) > radius: # continue distance = (i - startX) * (i - startX) + (j - startY) * (j - startY) if distance < ddradius: # 计算出（i,j）坐标的原坐标 # 计算公式中右边平方号里的部分 ratio = (ddradius - distance) / (ddradius - distance + ddmc) ratio = ratio * ratio # 映射原位置 UX = i - ratio * (endX - startX) UY = j - ratio * (endY - startY) # 根据双线性插值法得到UX，UY的值 # start_ = time.time() value = BilinearInsert(srcImg, UX, UY) # print(f"双线性插值耗时;{time.time() - start_}") # 改变当前 i ，j的值 copyImg[j + startTY, i + startTX] = value return copyImg # # 瘦脸pro2,采用了numpy自定义函数做处理 # def localTranslationWarpNumpy(self, srcImg, startP: np.matrix, endP: np.matrix, radius: float): # startX , startY = startP[0, 0], startP[0, 1] # endX, endY = endP[0, 0], endP[0, 1] # ddradius = float(radius * radius) # 圆的半径 # copyImg = srcImg.copy() # copy后的图像矩阵 # # 计算公式中的|m-c|^2 # ddmc = (endX - startX)**2 + (endY - startY)**2 # # 计算正方形的左上角起始点 # startTX, startTY = (startX - math.floor(radius + 1), startY - math.floor((radius + 1))) # # 计算正方形的右下角的结束点 # endTX, endTY = (startX + math.floor(radius + 1), startY + math.floor((radius + 1))) # # 剪切srcImg # self.thinImage = srcImg[startTY : endTY + 1, startTX : endTX + 1, :] # # s = self.thinImage # # db.cv_show(srcImg) # # 裁剪后的图像相当于在x,y都减少了startX - math.floor(radius + 1) # # 原本的endX, endY在切后的坐标点 # endX, endY = (endX - startX + math.floor(radius + 1), endY - startY + math.floor(radius + 1)) # # 原本的startX, startY剪切后的坐标点 # startX ,startY = (math.floor(radius + 1), math.floor(radius + 1)) # H, W, C = self.thinImage.shape # 获取图像的形状 # index_m = np.arange(H * W).reshape((H, W)) # triangle_ufunc = np.frompyfunc(self.process, 9, 3) # # start_ = time.time() # finalImgB, finalImgG, finalImgR = triangle_ufunc(index_m, self, W, ddradius, ddmc, startX, startY, endX, endY) # finaleImg = np.dstack((finalImgB, finalImgG, finalImgR)).astype(np.uint8) # finaleImg = np.fliplr(np.rot90(finaleImg, -1)) # copyImg[startTY: endTY + 1, startTX: endTX + 1, :] = finaleImg # # print(f"图像处理耗时;{time.time() - start_}") # # db.cv_show(copyImg) # return copyImg # 瘦脸pro3,采用opencv内置函数 @staticmethod def localTranslationWarpFastWithStrength(srcImg, startP: np.matrix, endP: np.matrix, radius, strength: float = 100.): """ 采用opencv内置函数 Args: srcImg: 源图像 startP: 起点位置 endP: 终点位置 radius: 处理半径 strength: 瘦脸强度，一般取100以上 Returns: """ startX, startY = startP[0, 0], startP[0, 1] endX, endY = endP[0, 0], endP[0, 1] ddradius = float(radius * radius) # copyImg = np.zeros(srcImg.shape, np.uint8) # copyImg = srcImg.copy() maskImg = np.zeros(srcImg.shape[:2], np.uint8) cv2.circle(maskImg, (startX, startY), math.ceil(radius), (255, 255, 255), -1) K0 = 100 / strength # 计算公式中的|m-c|^2 ddmc_x = (endX - startX) * (endX - startX) ddmc_y = (endY - startY) * (endY - startY) H, W, C = srcImg.shape mapX = np.vstack([np.arange(W).astype(np.float32).reshape(1, -1)] * H) mapY = np.hstack([np.arange(H).astype(np.float32).reshape(-1, 1)] * W) distance_x = (mapX - startX) * (mapX - startX) distance_y = (mapY - startY) * (mapY - startY) distance = distance_x + distance_y K1 = np.sqrt(distance) ratio_x = (ddradius - distance_x) / (ddradius - distance_x + K0 * ddmc_x) ratio_y = (ddradius - distance_y) / (ddradius - distance_y + K0 * ddmc_y) ratio_x = ratio_x * ratio_x ratio_y = ratio_y * ratio_y UX = mapX - ratio_x * (endX - startX) * (1 - K1 / radius) UY = mapY - ratio_y * (endY - startY) * (1 - K1 / radius) np.copyto(UX, mapX, where=maskImg == 0) np.copyto(UY, mapY, where=maskImg == 0) UX = UX.astype(np.float32) UY = UY.astype(np.float32) copyImg = cv2.remap(srcImg, UX, UY, interpolation=cv2.INTER_LINEAR) return copyImg def thinFace(src, landmark, place: int = 0, strength=30.): """ 瘦脸程序接口，输入人脸关键点信息和强度，即可实现瘦脸 注意处理四通道图像 Args: src: 原图 landmark: 关键点信息 place: 选择瘦脸区域，为0-4之间的值 strength: 瘦脸强度，输入值在0-10之间，如果小于或者等于0，则不瘦脸 Returns: 瘦脸后的图像 """ strength = min(100., strength * 10.) if strength <= 0.: return src # 也可以设置瘦脸区域 place = max(0, min(4, int(place))) left_landmark = landmark[4 + place] left_landmark_down = landmark[6 + place] right_landmark = landmark[13 + place] right_landmark_down = landmark[15 + place] endPt = landmark[58] # 计算第4个点到第6个点的距离作为瘦脸距离 r_left = math.sqrt( (left_landmark[0, 0] - left_landmark_down[0, 0]) ** 2 + (left_landmark[0, 1] - left_landmark_down[0, 1]) ** 2 ) # 计算第14个点到第16个点的距离作为瘦脸距离 r_right = math.sqrt((right_landmark[0, 0] - right_landmark_down[0, 0]) ** 2 + (right_landmark[0, 1] - right_landmark_down[0, 1]) ** 2) # 瘦左边脸 thin_image = TranslationWarp.localTranslationWarpFastWithStrength(src, left_landmark[0], endPt[0], r_left, strength) # 瘦右边脸 thin_image = TranslationWarp.localTranslationWarpFastWithStrength(thin_image, right_landmark[0], endPt[0], r_right, strength) return thin_image if __name__ == "__main__": import os from hycv.FaceDetection68.faceDetection68 import FaceDetection68 local_file = os.path.dirname(__file__) PREDICTOR_PATH = f"{local_file}/weights/shape_predictor_68_face_landmarks.dat" # 关键点检测模型路径 fd68 = FaceDetection68(model_path=PREDICTOR_PATH) input_image = cv2.imread("test_image/4.jpg", -1) _, landmark_, _ = fd68.facePoints(input_image) output_image = thinFace(input_image, landmark_, strength=30.2) cv2.imwrite("thinFaceCompare.png", np.hstack((input_image, output_image)))