用Retinaface_pytorch项目检测人脸+ Python 3 剪切人脸图片并保存

概述

因为我的项目中缺一不可的一部分是提取人脸特征，因为以前并没有接触过人脸的项目，所以想把自己每次的一点点进步和学习都记录下来，这次是记录我学习切割人脸图片并保存，以及将人脸检测框和五个特征点坐标信息保存在.txt文件中。
这篇文档的方法一要稍微复杂一点点，我结合Retinaface_pytorch项目中的detect.py（检测单张图片）来检测人脸，并通过以下方法一成功的剪切出人脸，并保存特征点信息。方法二要更简单一些，相当于我用一行代码代替了方法二中的多行代码，并实现了批量检测并切割人脸图片并保存。

首先附上Retinaface_pytorch项目的GitHub链接：

https://github.com/biubug6/Pytorch_Retinaface

一. 复现Retinaface_pytorch

根据Retina_pytorch项目在GitHub上给出的教程来复现Retinaface_pytorch项目的预训练模型。跑通项目中的detect.py，改变程序中的图片路径你可以试一下检测自己的单张照片。确保detect.py可以正常检测人脸。

运行效果图：

二.剪切人脸图（方法一）

1. 方法简述

参考：Python 3 利用 Dlib 实现人脸检测和剪切
这一篇博客介绍了利用 Python 开发，借助 Dlib 库进行人脸检测 / face detection 和剪切。里面既讲了将检测到的人脸剪切下来，依次排序平铺显示在新的图像上，也讲了将检测到的人脸存储为单个人脸图像，讲的很清晰明了。主要的大概思路就是根据检测到的人脸框坐标，计算人脸框的大小，根据这个人脸框大小生成空白的图片，接着再将人脸的图的像素点依次填充在空白图上，完成人脸的剪切。

将检测到的人脸剪切下来，依次排序平铺显示在新的图像上（效果图）
将检测到的人脸存储为单个人脸图像（效果图）

Retinaface_pytorch的detect.py检测单张图片中的人脸+剪切单个人脸图

#调用相关模块和包
from __future__ import print_function
import os
import argparse
import torch
import torch.backends.cudnn as cudnn
import numpy as np
from data import cfg_mnet, cfg_re50
from layers.functions.prior_box import PriorBox
from utils.nms.py_cpu_nms import py_cpu_nms
import cv2
from models.retinaface import RetinaFace
from utils.box_utils import decode, decode_landm
import timeparser = argparse.ArgumentParser(description='Retinaface')
#设置供选择的参数（基本上保持Retinaface_pytorch中detect.py的代码不变添加或改动了几个参数使得该项目默认在CPU上运行）
parser.add_argument('-m', '--trained_model', default='./weights/Resnet50_Final.pth',type=str, help='Trained state_dict file path to open')
parser.add_argument('--network', default='resnet50', help='Backbone network mobile0.25 or resnet50')
parser.add_argument('--cpu', action="store_true", default=True, help='Use cpu inference')
parser.add_argument('--confidence_threshold', default=0.02, type=float, help='confidence_threshold')
parser.add_argument('--top_k', default=5000, type=int, help='top_k')
parser.add_argument('--nms_threshold', default=0.4, type=float, help='nms_threshold')
parser.add_argument('--keep_top_k', default=750, type=int, help='keep_top_k')
parser.add_argument('-s', '--show_image', action="store_true", default=False, help='show detection results')
parser.add_argument('--vis_thres', default=0.6, type=float, help='visualization_threshold')
parser.add_argument('--show_cutting_image', action ="store_true", default =True, help = 'show_crop_images')
parser.add_argument('--save_folder', default='curve/', type=str, help='Dir to save results')
args = parser.parse_args()#保持源代码不变
def check_keys(model, pretrained_state_dict):ckpt_keys = set(pretrained_state_dict.keys())model_keys = set(model.state_dict().keys())used_pretrained_keys = model_keys & ckpt_keysunused_pretrained_keys = ckpt_keys - model_keysmissing_keys = model_keys - ckpt_keysprint('Missing keys:{}'.format(len(missing_keys)))print('Unused checkpoint keys:{}'.format(len(unused_pretrained_keys)))print('Used keys:{}'.format(len(used_pretrained_keys)))assert len(used_pretrained_keys) > 0, 'load NONE from pretrained checkpoint'return True#保持源代码不变
def remove_prefix(state_dict, prefix):''' Old style model is stored with all names of parameters sharing common prefix 'module.' '''print('remove prefix \'{}\''.format(prefix))f = lambda x: x.split(prefix, 1)[-1] if x.startswith(prefix) else xreturn {f(key): value for key, value in state_dict.items()}#保持源代码不变
def load_model(model, pretrained_path, load_to_cpu):print('Loading pretrained model from {}'.format(pretrained_path))if load_to_cpu:pretrained_dict = torch.load(pretrained_path, map_location=lambda storage, loc: storage)else:device = torch.cuda.current_device()pretrained_dict = torch.load(pretrained_path, map_location=lambda storage, loc: storage.cuda(device))if "state_dict" in pretrained_dict.keys():pretrained_dict = remove_prefix(pretrained_dict['state_dict'], 'module.')else:pretrained_dict = remove_prefix(pretrained_dict, 'module.')check_keys(model, pretrained_dict)model.load_state_dict(pretrained_dict, strict=False)return model#保持源代码不变
if __name__ == '__main__':torch.set_grad_enabled(False)cfg = Noneif args.network == "mobile0.25":cfg = cfg_mnetelif args.network == "resnet50":cfg = cfg_re50# net and modelnet = RetinaFace(cfg=cfg, phase = 'test')net = load_model(net, args.trained_model, args.cpu)net.eval()print('Finished loading model!')print(net)cudnn.benchmark = True #提高运行效率device = torch.device("cpu" if args.cpu else "cuda")net = net.to(device)resize = 1# testing beginfor i in range(100):image_path = "./curve/test.jpg" #将这里的图片路径换成你自己的图片路径img_raw = cv2.imread(image_path, cv2.IMREAD_COLOR) #读取彩色图片img = np.float32(img_raw)im_height, im_width, _ = img.shapescale = torch.Tensor([img.shape[1], img.shape[0], img.shape[1], img.shape[0]])img -= (104, 117, 123)img = img.transpose(2, 0, 1)  #对图片进行转置处理img = torch.from_numpy(img).unsqueeze(0)img = img.to(device)scale = scale.to(device)tic = time.time() #计时loc, conf, landms = net(img)  # forward passprint('net forward time: {:.4f}'.format(time.time() - tic))priorbox = PriorBox(cfg, image_size=(im_height, im_width))priors = priorbox.forward()priors = priors.to(device)prior_data = priors.databoxes = decode(loc.data.squeeze(0), prior_data, cfg['variance'])boxes = boxes * scale / resizeboxes = boxes.cpu().numpy()scores = conf.squeeze(0).data.cpu().numpy()[:, 1]landms = decode_landm(landms.data.squeeze(0), prior_data, cfg['variance'])scale1 = torch.Tensor([img.shape[3], img.shape[2], img.shape[3], img.shape[2],img.shape[3], img.shape[2], img.shape[3], img.shape[2],img.shape[3], img.shape[2]])scale1 = scale1.to(device)landms = landms * scale1 / resizelandms = landms.cpu().numpy()# ignore low scoresinds = np.where(scores > args.confidence_threshold)[0]boxes = boxes[inds]landms = landms[inds]scores = scores[inds]# keep top-K before NMSorder = scores.argsort()[::-1][:args.top_k]boxes = boxes[order]landms = landms[order]scores = scores[order]# do NMSdets = np.hstack((boxes, scores[:, np.newaxis])).astype(np.float32, copy=False)keep = py_cpu_nms(dets, args.nms_threshold)# keep = nms(dets, args.nms_threshold,force_cpu=args.cpu)dets = dets[keep, :]landms = landms[keep]# keep top-K faster NMSdets = dets[:args.keep_top_k, :]landms = landms[:args.keep_top_k, :]dets = np.concatenate((dets, landms), axis=1)
------------------------------------------------------------  #到这里为止我们已经利用Retinaface_pytorch的预训练模型检测完了人脸，并获得了人脸框和人脸五个特征点的坐标信息，全保存在dets中，接下来为人脸剪切部分# 用来储存生成的单张人脸的路径path_save = "./curve/faces/" #你可以将这里的路径换成你自己的路径#剪切图片if args.show_cutting_image：for num, b in enumerate(dets): # dets中包含了人脸框和五个特征点的坐标if b[4] < args.vis_thres:continueb = list(map(int, b))# landms，在人脸上画出特征点，要是你想保存不显示特征点的人脸图，你可以把这里注释掉cv2.circle(img_raw, (b[5], b[6]), 1, (0, 0, 255), 4)cv2.circle(img_raw, (b[7], b[8]), 1, (0, 255, 255), 4)cv2.circle(img_raw, (b[9], b[10]), 1, (255, 0, 255), 4)cv2.circle(img_raw, (b[11], b[12]), 1, (0, 255, 0), 4)cv2.circle(img_raw, (b[13], b[14]), 1, (255, 0, 0), 4)#计算人脸框矩形大小Height = b[3] - b[1]Width = b[2] - b[0]# 显示人脸矩阵大小print("人脸数 / faces in all:", str(num+1), "\n")print("窗口大小 / The size of window:", '\n', "高度 / height:", Height, '\n', "宽度 / width: ", Width)#根据人脸框大小，生成空白的图片img_blank = np.zeros((Height, Width, 3), np.uint8)# 将人脸填充到空白图片for h in range(Height):for w in range(Width):img_blank[h][w] = img_raw[b[1] + h][b[0] + w]cv2.namedWindow("img_faces")  # , 2)cv2.imshow("img_faces", img_blank)  #显示图片cv2.imwrite(path_save + "img_face_4" + str(num + 1) + ".jpg", img_blank)  #将图片保存至你指定的文件夹print("Save into:", path_save + "img_face_4" + str(num + 1) + ".jpg")cv2.waitKey(0)#保存人脸框，特征点的坐标信息到txt中if not os.path.exists(args.save_folder):os.makedirs(args.save_folder)fw = open(os.path.join(args.save_folder, '__dets.txt'), 'w')  #在指定的文件夹中生成并打开一个名为__dets的txt文件if args.save_folder:fw.write('{:s}\n'.format(img_name)) #在txt中写入图片名字for k in range(dets.shape[0]): #遍历dets中的坐标信息，dets中的信息包括（人脸框的左上角x,y坐标，人脸框右下角x,y坐标，人脸检测精度scores，五个特征点的x,y坐标，共15个信息）xmin = dets[k, 0]ymin = dets[k, 1]xmax = dets[k, 2]ymax = dets[k, 3]score = dets[k, 4]w = xmax - xmin + 1h = ymax - ymin + 1landms1_x = dets[k, 5]landms1_y = dets[k, 6]landms2_x = dets[k, 7]landms2_y = dets[k, 8]landms3_x = dets[k, 9]landms3_y = dets[k, 10]landms4_x = dets[k, 11]landms4_y = dets[k, 12]landms5_x = dets[k, 13]landms5_y = dets[k, 14]#将人脸框，人脸检测精度，五个特征点的坐标信息写入到txt文件中fw.write('{:d} {:d} {:d} {:d} {:.10f} {:d} {:d} {:d} {:d} {:d} {:d} {:d} {:d} {:d} {:d}\n'.format(int(xmin), int(ymin), int(w), int(h), score, int(landms1_x),int(landms1_y),int(landms2_x), int(landms2_y), int(landms3_x), int(landms3_y), int(landms4_x),int(landms4_y), int(landms5_x), int(landms5_y)))#写入完毕，关闭txt文件fw.close()

通过结合Retinaface_pytorch的detect.py代码和**参考：Python 3 利用 Dlib 实现人脸检测和剪切的代码，我完成了在单张图片中检测人脸的任务。但是我的初衷是想完成批量检测人脸，切割人脸图片并保存，在我想继续将参考：Python 3 利用 Dlib 实现人脸检测和剪切**的剪切人脸的代码和Retina_pytorch的test_fddb.py的代码相结合来批量检测人脸时，总是会报错IndexError: index 368 is out of bounds for axis 0 with size 368，找了好久也找不到解决办法，所以干脆听坤神的换了一种剪切人脸图的简单方法，一行代码就搞定了。如下方法二所示。

三、剪切人脸（方法二）—批量剪切单个人脸图并保存

将上面代码中剪切人脸的这一部分

               #计算人脸框矩形大小Height = b[3] - b[1]Width = b[2] - b[0]# 显示人脸矩阵大小print("人脸数 / faces in all:", str(num+1), "\n")print("窗口大小 / The size of window:", '\n', "高度 / height:", Height, '\n', "宽度 / width: ", Width)#根据人脸框大小，生成空白的图片img_blank = np.zeros((Height, Width, 3), np.uint8)# 将人脸填充到空白图片for h in range(Height):for w in range(Width):img_blank[h][w] = img_raw[b[1] + h][b[0] + w]

全部用下面这一句代码来代替

img_blank = img_raw[int(b[1]):int(b[3]), int(b[0]):int(b[2])] # height, width

然后你再将人脸剪切部分的代码放在Retinaface_pytorch的test_fddb.py代码的后面，你就可以开始尝试检测批量剪切人脸的效果了，即一次性把一个数据集中的所有图片的人脸都检测，切割，并将每个人脸保存为单张图片，效果如图所示。

本人是初学人脸的小白，只是想利用博客来记录并总结自己的学习进度，如果能帮到同样刚开始学人脸的小白，我会很开心。若有更好的方法，或不太准确的地方，欢迎大家一起讨论指正。