基于OpenCV与Keras深度学习网络的活体检测的代码及训练的模型.zip

# USAGE # python liveness_demo.py --model liveness.model --detector face_detector # from imutils.video import VideoStream from keras.preprocessing.image import img_to_array from keras.models import load_model import numpy as np import argparse import cv2 import os import time # 参数设置 ap = argparse.ArgumentParser() ap.add_argument("-m", "--model", type=str, required=True, help="path to trained model") ap.add_argument("-d", "--detector", type=str, required=True, help="path to OpenCV's deep learning face detector") ap.add_argument("-c", "--confidence", type=float, default=0.6, help="minimum probability to filter weak detections") args = vars(ap.parse_args()) # 判断活体参数初始化 count = 0 # 人脸计数 flag = 0 timeline = 0 # 时间计数 sign = 0 # 状态信息 label_list = [] # 人脸检测器 print("[INFO] loading face detector...") protoPath = os.path.sep.join([args["detector"], "deploy.prototxt"]) modelPath = os.path.sep.join([args["detector"], "res10_300x300_ssd_iter_140000.caffemodel"]) net = cv2.dnn.readNetFromCaffe(protoPath, modelPath) # 加载模型 print("[INFO] loading liveness detector...") model = load_model(args["model"]) labels = ["Fake", "Real"] print("[INFO] starting video stream...") # vs = VideoStream(src="test.mp4").start() # vs = VideoStream(“test.mp4”).start() vs = cv2.VideoCapture(1) # size 被按比例压缩到最大width为600 r = float(int(vs.get(cv2.CAP_PROP_FRAME_HEIGHT)) / int(vs.get(cv2.CAP_PROP_FRAME_WIDTH))) size = (600, int(600 * r)) # size = (int(vs.get(cv2.CAP_PROP_FRAME_WIDTH)), int(vs.get(cv2.CAP_PROP_FRAME_HEIGHT))) writer = cv2.VideoWriter("model6_out3.mp4", cv2.VideoWriter_fourcc('m', 'p', '4', 'v'), 15.0, size) print("(width, height)", size) # count = fake_num = real_num = 0 # judge = "" while True: ret, frame = vs.read() if not ret: continue # else: # count +=1 # # # 10帧为窗口判断真假 # if count % 11 ==0: # count = 0 # elif count is 10:+ # if fake_num >= 9: # judge = "fake" # if real_num >= 9: # judge = "real" # real_num = 0 # fake_num = 0 # print("count ",count, "fake_num ", fake_num, "real_num ", real_num, "judge ", judge) # size 被按比例压缩到最大width为600 t1 = time.time() frame = cv2.resize(frame, size) # opencv dnn 模块预处理，提取blob (h, w) = frame.shape[:2] blob = cv2.dnn.blobFromImage(cv2.resize(frame, (300, 300)), 1.0, (300, 300), (104.0, 177.0, 123.0)) # 前向传播，得到预测结果 net.setInput(blob) detections = net.forward() # 检测到人脸时 for i in range(0, detections.shape[2]): # 置信度 confidence = detections[0, 0, i, 2] # 过滤掉低置信度人脸 # print(confidence) if confidence > args["confidence"]: # 得到人脸roi box = detections[0, 0, i, 3:7] * np.array([w, h, w, h]) (startX, startY, endX, endY) = box.astype("int") CenterX = (startX + endX) / 2 CenterY = (startY + endY) / 2 face_width = endX - startX face_height = endY - startY face_width = face_width * 1.5 face_height = face_height * 1.5 big_startX = int(CenterX - (face_width / 2)) big_endX = int(CenterX + (face_width / 2)) big_startY = int(CenterY - (face_height / 2)) big_endY = int(CenterY + (face_height / 2)) # 确保roi 小于帧的尺寸范围 if big_startX < 0 or big_startY < 0 or big_endX > w or big_endY > h: continue # frame[startY:endY, startX:endX] = (0,0,0) # resize到32*32，与神经网络的输入尺寸一致 face = frame[big_startY:big_endY, big_startX:big_endX] if face.size == 0: continue face = cv2.resize(face, (96, 96)) face = face.astype("float") / 255.0 face = img_to_array(face) face = np.expand_dims(face, axis=0) # 前向传播liveness.model 得到预测结果 preds = model.predict(face)[0] j = np.argmax(preds) # print(preds, j) label = labels[j] # print("label", label, "j", type(j)) # j = int(j) # if j is 0: # fake_num += 1 # elif j is 1: # real_num +=1 # 检验到人脸则count计数加一 if j == 1: count += 1 # 输出结果 # 每8帧是一个计时单位，若8帧里面超过6帧为真则返回真 if timeline == 8: # label = "{}: {:.4f}".format(label, preds[j]) if count >= 6: label = 'Real' else: label = 'Fake' label_list.append(label) count = 0 timeline = 0 # 3个计数单位后进行判断 # 8*3 = 24 # 这样可以降低错误率一丢丢，另一种思路就是24帧中有18帧即可 if len(label_list) == 3: if label_list.count('Real') >= label_list.count('Fake'): label_new = 'Real' else: label_new = 'Fake' del label_list[:] # 清空label_list缓存 sign = 1 # 判断完成的标志 t2 = time.time() print(t2-t1, label_new) # return label_new # 判断完成后对判断结果进行分析，Real则返回绿色框框，Fake则返回红色框框 # print(sign) if sign == 1: label = "{}".format(label_new) # print(label) if label == 'Real': cv2.putText(frame, label_new, (startX, startY - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2) cv2.rectangle(frame, (startX, startY), (endX, endY), (0, 255, 0), 2) else: cv2.putText(frame, label_new, (startX, startY - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2) cv2.rectangle(frame, (startX, startY), (endX, endY), (0, 0, 255), 2) # cache += 1 # if cache % 52 == 0: # sign = 0 # 若判断为完成，则继续计数并显示loading框，直至计数完成 # Usage: 若发现loading后面的小点卡住不变了，可能需要调整下位置。 # 若识别对象的位置状态没有改变的话，Real或者Fake框会保持不变 否则会重新加载 elif sign == 0: timeline += 1 t = int(timeline / 2) + 1 loading = "loading{}".format('.' * t) label = loading cv2.putText(frame, label, (startX, startY - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2) cv2.rectangle(frame, (startX, startY), (endX, endY), (0, 0, 255), 2) # 摄像头连续8帧未捕捉到人则初始化窗口 # print(confidence) # 单人版，当摄像头里面出现小脸，即便无法识别真假也不会触发此条 # 无法解决的BUG：demo工作原理是一次对截取的每一帧的所有图像进行处理，然后再读取下一帧处理，无法将两个人脸分开 # 因此如果同时出现两张人脸，只会出现一种相同的结果 # print((confidence <= 0.072) or (confidence >= 1) if (confidence <= 0.072) or (confidence >= 0.12): # print('没发现人！！') flag = flag + 1 if flag >= 8: sign = 0 timeline = 0 count