DY23

flyskynet · Jul 1, 2022 · ea2b357 · ea2b357
1 parent 2fcf1b5
commit ea2b357
Show file tree

Hide file tree

Showing 7 changed files with 692 additions and 0 deletions.
diff --git a/.DS_Store b/.DS_Store
diff --git a/README.md b/README.md
@@ -21,6 +21,7 @@
 
 | 名称与在线播放地址                                           | 截图                                                         | 方法                                     | 难度 | 代码地址                                                     |
 | ------------------------------------------------------------ | ------------------------------------------------------------ | ---------------------------------------- | ---- | ------------------------------------------------------------ |
+| 23.facenet 戴口罩的人脸识别                                  | <img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/imgIMG_7913.PNG?x-oss-process=style/wp" style="width:200px;" /> | facenet、inception resnet、人脸识别      | ⭐️⭐️⭐️⭐️ | [codes/23.facenet戴口罩的人脸识别](https://github.com/enpeizhao/CVprojects/tree/main/codes/23.facenet戴口罩的人脸识别) |
 | 22.openvino树莓派扑克牌检测                                  | <img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img20220617173131.png?x-oss-process=style/wp" style="width:200px;" /> | Openvino、YOLOV5、树莓派                 | ⭐️⭐️   | [codes/22.openvino扑克牌检测](https://github.com/enpeizhao/CVprojects/tree/main/codes/22.openvino扑克牌检测) |
 | 21.GAN老照片上色并动起来                                     | <img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/imgIMG_7673.jpg?x-oss-process=style/wp" style="width:200px;" /> | GAN风格化、超分辨率、first order model   | ⭐️⭐️   | [codes/21.GAN老照片上色动起来](https://github.com/enpeizhao/CVprojects/tree/main/codes/21.GAN老照片上色动起来) |
 | 20.骨骼点动态动作识别                                        | <img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img20220518201724.png?x-oss-process=style/wp" style="width:200px;" /> | 姿态估计、DTW                            | ⭐️⭐️⭐️  | [codes/20.骨骼点动作识别](https://github.com/enpeizhao/CVprojects/tree/main/codes/20.骨骼点动作识别) |

diff --git a/codes/.DS_Store b/codes/.DS_Store
diff --git a/codes/23.facenet戴口罩的人脸识别/.DS_Store b/codes/23.facenet戴口罩的人脸识别/.DS_Store
diff --git a/codes/23.facenet戴口罩的人脸识别/README.md b/codes/23.facenet戴口罩的人脸识别/README.md
@@ -0,0 +1,55 @@
+# facenet戴口罩的人脸识别
+
+![](https://enpei-md.oss-cn-hangzhou.aliyuncs.com/imgIMG_7913.PNG?x-oss-process=style/wp)
+
+
+
+## 说明：
+
+* 本项目是本人为RESNET、inception授课做的应用演示。
+
+
+## 一、硬件：
+
+* Windows 
+* macOS
+
+## 二、软件环境安装：
+
+* TensorFlow 2
+* pytorch 1.10
+* OpenCV
+
+## 三、用法：
+
+* 下载压缩包`23.facenet_mask_recognition.zip`解压到`demo_mac.py`同级目录：[下载地址](https://github.com/enpeizhao/CVprojects/releases)
+
+  <img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img20220701105256.png?x-oss-process=style/wp" style="zoom:50%;" />
+
+* 人脸库照片放到`images/origin`下，命名规则为`人名_序号.jpg`，如`恩培_1.jpg`，`恩培_2.jpg`，一个人建议放个1~3张清晰度较好的图片
+
+<img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img20220701110037.png?x-oss-process=style/wp" style="zoom: 50%;" />
+
+* Windows 运行：`python demo_windows.py --threshold={识别阈值}`如`python demo_windows.py --threshold=1`，阈值越低要求越严格
+* macOS运行：`python demo_mac.py --threshold={识别阈值}`
+
+## 四、其他信息：
+
+* 模型架构：inception_resnet_v1
+* 训练集：1.4W人80W张图片（一半是合成的戴口罩数据集）
+
+<img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img20220701105743.png?x-oss-process=style/wp" style="zoom: 25%;" />
+
+* 测试ACC：1.2W张测试样本上，ACC最高为97.1%；
+* 训练时间：A100 145EPOCH，28个小时
+* ![](https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img1.png?x-oss-process=style/wp)
+* ![](https://enpei-md.oss-cn-hangzhou.aliyuncs.com/img2.png?x-oss-process=style/wp)
+
+
+
+
+
+### 微信技术交流、问题反馈：
+
+<img src="https://enpei-md.oss-cn-hangzhou.aliyuncs.com/imgIMG_5862.JPG?x-oss-process=style/wp" style="width:200px;" />
+
diff --git a/codes/23.facenet戴口罩的人脸识别/demo_mac.py b/codes/23.facenet戴口罩的人脸识别/demo_mac.py
@@ -0,0 +1,319 @@
+from model.SSD import FaceMaskDetection
+from model.FACENET import InceptionResnetV1
+import torch
+
+import cv2
+import numpy as np
+import glob
+import time
+import tqdm
+from PIL import Image,ImageDraw,ImageFont
+import os
+from argparse import ArgumentParser
+
+class MaskedFaceRecog:
+    def __init__(self):
+        # 加载检测模型
+        face_mask_model_path = r'weights/SSD/face_mask_detection.pb'
+        self.ssd_detector = FaceMaskDetection(face_mask_model_path,margin=0,GPU_ratio=0.1)
+
+
+        # 加载识别模型
+        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        # 实例化
+        self.facenet = InceptionResnetV1(is_train=False,embedding_length=128,num_classes=14575).to(self.device)
+        # 从训练文件中加载
+        self.facenet.load_state_dict(torch.load(r'./weights/4.Chinese_CASIA_ALL_AG_epoch/facenet_best.pt',map_location=self.device))
+        self.facenet.eval()
+
+
+        # 加载目标人的特征
+        # name_list支持：恩培,恩培_1,,恩培_2形式，与known_embedding对应
+        self.name_list,self.known_embedding = self.loadFaceFeats()
+        # 增加一个未知人员
+        self.name_list.append('未知')
+        # 生成每个人的名称PNG图片（以解决中文显示问题）
+        self.name_png_list = self.getNamePngs(self.name_list)
+        # 加载佩戴和未佩戴标志
+        self.mask_class_overlay = self.getMaskClassPngs()
+
+    def getMaskClassPngs(self):
+        '''
+        加载佩戴和未佩戴标志
+        '''
+        labels = ['masked','without_mask']
+        overlay_list = []
+        for label in labels:
+            fileName = './images/%s.png' % (label)
+            overlay = cv2.imread(fileName,cv2.COLOR_RGB2BGR)
+            overlay = cv2.resize(overlay,(0,0), fx=0.2, fy=0.2)
+            overlay_list.append(overlay)
+        return overlay_list
+
+    def readPngFile(self,fileName):
+        '''
+        读取PNG图片
+        '''
+        # 解决中文路径问题
+        png_img = cv2.imdecode(np.fromfile(fileName,dtype=np.uint8),-1)
+        # 转为BGR，变成3通道
+        png_img = cv2.cvtColor(png_img,cv2.COLOR_RGB2BGR)
+        png_img = cv2.resize(png_img,(0,0), fx=0.4, fy=0.4)
+        return png_img
+
+
+    def getNamePngs(self,name_list):
+        '''
+        生成每个人的名称PNG图片（以解决中文显示问题）
+        '''
+        # 先将['恩培','恩培_1','恩培_2','小明','小明_1','小明_2']变成['恩培','小明']
+        real_name_list = []
+        for name in name_list:
+            real_name = name.split('_')[0]
+            if real_name not in real_name_list:
+                real_name_list.append(real_name)
+
+
+        pngs_list = {}
+        for name in tqdm.tqdm(real_name_list,desc='生成人脸标签PNG...'):
+
+
+            filename = './images/name_png/'+name+'.png'
+            # 如果存在，直接读取
+            if os.path.exists(filename):
+                png_img = self.readPngFile(filename)
+                pngs_list[name] = png_img
+                continue
+
+            # 如果不存在，先生成
+            # 背景
+            bg = Image.new("RGBA",(400,100),(0,0,0,0))
+            # 添加文字
+            d = ImageDraw.Draw(bg)
+            font  = ImageFont.truetype('./fonts/MSYH.ttc',80,encoding="utf-8")
+
+            if name == '未知':
+                color = (0,0,255,255)
+            else:
+                color = (0,255,0,255)
+
+            d.text((0,0),name,font=font,fill=color)
+            # 保存
+            bg.save(filename)
+            # 再次检查
+            if os.path.exists(filename):
+                png_img = self.readPngFile(filename)
+                pngs_list[name] = png_img
+
+        return pngs_list
+
+    def loadFaceFeats(self):
+        '''
+        加载目标人的特征
+        '''
+        # 记录名字
+        name_list = []
+        # 输入网络的所有人脸图片
+        known_faces_input = []
+        # 遍历
+        known_face_list = glob.glob('./images/origin/*')
+        for face in tqdm.tqdm(known_face_list,desc='处理目标人脸...'):
+            name = face.split('/')[-1].split('.')[0]
+            name_list.append(name)
+            # 裁剪人脸
+            croped_face = self.getCropedFaceFromFile(face)
+            if croped_face is None:
+                print('图片：{} 未检测到人脸，跳过'.format(face))
+                continue
+            # 预处理
+            img_input = self.imgPreprocess(croped_face)
+            known_faces_input.append(img_input)
+        # 转为Nummpy
+        faces_input = np.array(known_faces_input)
+        # 转tensor并放到GPU
+        tensor_input = torch.from_numpy(faces_input).to(self.device)
+        # 得到所有的embedding,转numpy
+        known_embedding = self.facenet(tensor_input).detach().cpu().numpy()
+
+        return name_list,known_embedding
+
+    def getCropedFaceFromFile(self,img_file, conf_thresh=0.5 ):
+
+        # 读取图片
+        # 解决中文路径问题
+        img_ori = cv2.imdecode(np.fromfile(img_file,dtype=np.uint8),-1)
+
+        if img_ori is None:
+            return None
+        # 转RGB
+        img = cv2.cvtColor(img_ori,cv2.COLOR_BGR2RGB)
+        # 缩放
+        img = cv2.resize(img,self.ssd_detector.img_size)
+        # 转float32
+        img = img.astype(np.float32)
+        # 归一
+        img /= 255
+        # 增加维度
+        img_4d = np.expand_dims(img,axis=0)
+        # 原始高度和宽度
+        ori_h,ori_w = img_ori.shape[:2]
+        bboxes, re_confidence, re_classes, re_mask_id = self.ssd_detector.inference(img_4d,ori_h,ori_w)
+        for index,bbox in enumerate(bboxes):
+            class_id = re_mask_id[index] 
+            l,t,r,b = bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]
+
+            croped_face = img_ori[t:b,l:r]
+            return croped_face
+
+        # 都不满足
+        return None
+
+    def imgPreprocess(self,img):
+        # 转为float32
+        img = img.astype(np.float32)
+        # 缩放
+        img = cv2.resize(img,(112,112))
+        # BGR 2 RGB
+        img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
+        # h,w,c 2 c,h,w
+        img = img.transpose((2,0,1))
+        # 归一化[0,255] 转 [-1,1]
+        img = (img - 127.5) / 127.5
+        # 增加维度
+        # img = np.expand_dims(img,0)
+
+        return img
+
+    def main(self,threshold):
+
+        cap = cv2.VideoCapture(0)
+        frame_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        frame_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+
+        png_index = 0
+
+        while True:
+            start_time = time.time()
+
+            ret,frame = cap.read()
+            frame = cv2.flip(frame,1)
+            # 转RGB
+            img = cv2.cvtColor(frame,cv2.COLOR_BGR2RGB)
+            # 缩放
+            img = cv2.resize(img,self.ssd_detector.img_size)
+            # 转float32
+            img = img.astype(np.float32)
+            # 归一
+            img /= 255
+            # 增加维度
+            img_4d = np.expand_dims(img,axis=0)
+            bboxes, re_confidence, re_classes, re_mask_id = self.ssd_detector.inference(img_4d,frame_h,frame_w)
+
+            for index,bbox in enumerate(bboxes):
+                class_id = re_mask_id[index] 
+                conf  = re_confidence[index] 
+
+                if class_id == 0:
+                    color = (0, 255, 0)  # 戴口罩
+                elif class_id == 1:
+                    color = (0, 0, 255)  # 没带口罩
+
+
+
+                l,t,r,b = bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]
+
+                # cv2.putText(frame,str(round(conf,2)),(l,t-10),cv2.FONT_ITALIC,1,(0,255,0),1)
+
+                # 裁剪人脸
+                crop_face = frame[t:b,l:r]
+
+
+                # 人脸识别
+
+                # 转为float32
+                img = crop_face.astype(np.float32)
+                # 缩放
+                img = cv2.resize(img,(112,112))
+                # BGR 2 RGB
+                img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
+                # h,w,c 2 c,h,w
+                img = img.transpose((2,0,1))
+                # 归一化[0,255] 转 [-1,1]
+                img = (img - 127.5) / 127.5
+                # 扩展维度
+                img_input = np.expand_dims(img,0)
+                # C连续特性
+                # img_input = np.ascontiguousarray(img_input)
+                # 转tensor并放到GPU
+                tensor_input = torch.from_numpy(img_input).to(self.device)
+                # 得到embedding
+                embedding = self.facenet(tensor_input)
+                embedding = embedding.detach().cpu().numpy()
+                # print(embedding)
+                # 计算距离
+                dist_list = np.linalg.norm((embedding-self.known_embedding),axis=1)
+                # 最小距离索引
+                min_index = np.argmin(dist_list)
+                # 识别人名与距离
+                pred_name = self.name_list[min_index]
+                # 最短距离
+                min_dist = dist_list[min_index]
+
+                if min_dist < threshold:
+                    # 识别到人
+                    # 人名png
+                    real_name = pred_name.split('_')[0]
+                    name_overlay = self.name_png_list[real_name]
+                else:
+                    # 未识别到，加载未知
+                    name_overlay = self.name_png_list['未知']
+
+                # √和×标志
+                class_overlay = self.mask_class_overlay[class_id]
+
+
+                # 拼接两个PNG
+                overlay = np.zeros((40, 210,3), np.uint8)
+                overlay[:40, :40] = class_overlay
+                overlay[:40, 50:210] = name_overlay
+
+
+                # 覆盖显示
+                overlay_h,overlay_w = overlay.shape[:2]
+                # 覆盖范围
+                overlay_l,overlay_t = l,(t - overlay_h-20)
+                overlay_r,overlay_b = (l + overlay_w),(overlay_t+overlay_h)
+                # 判断边界
+                if overlay_t > 0 and overlay_r < frame_w:
+                    overlay_copy=cv2.addWeighted(frame[overlay_t:overlay_b, overlay_l:overlay_r ],1,overlay,20,0)
+
+                    frame[overlay_t:overlay_b, overlay_l:overlay_r ] =  overlay_copy
+
+                print(pred_name,min_dist)
+
+
+                cv2.rectangle(frame,(l,t),(r,b),color,2)
+
+
+            fps = 1/ (time.time()- start_time)
+            cv2.putText(frame,str(round(fps,2)),(50,50),cv2.FONT_ITALIC,1,(0,255,0),2)
+
+            cv2.imshow('demo',frame)
+            if cv2.waitKey(1) & 0xff == ord('q'):
+                break
+        cap.release()
+        cv2.destroyAllWindows()
+
+
+
+
+
+# 参数
+parser = ArgumentParser()
+parser.add_argument("--threshold", type=float, default=1,
+                    help="人脸识别距离阈值，越低越准确")                                         
+
+args = parser.parse_args()
+
+mf = MaskedFaceRecog()
+mf.main(threshold=args.threshold)