OpenCV头像训练与识别

识别效果图

依赖安装 opencv-python opencv-contrib-python tensorflow Pillow numpy 

pip3 install opencv-python&＃61;&＃61;4.5.1.48 -i https://pypi.tuna.tsinghua.edu.cn/simple --default-timeout&＃61;10000
pip3 install opencv-contrib-python&＃61;&＃61;4.5.1.48 -i https://pypi.tuna.tsinghua.edu.cn/simple --default-timeout&＃61;10000
pip3 install tensorflow&＃61;&＃61;2.4.1 -i https://pypi.tuna.tsinghua.edu.cn/simple --default-timeout&＃61;10000
pip3 install -U pillow&＃61;&＃61;8.1.2
pip3 install numpy&＃61;&＃61;1.19.4 -i https://pypi.tuna.tsinghua.edu.cn/simple --default-timeout&＃61;10000

头像目标锁定资源文件 haarcascade_frontalface_default.xml  如果觉得慢,可以从 csdn 资源中下载 

项目结构

采集数据 collection.py  采集完成之后可以通过肉眼删除误差较大的数据

# 头像数据采集
import cv2 as cv
import uuid# 数据集收集
def split_image(frame, set_path):# 图片转灰度frame_gray &＃61; cv.cvtColor(frame, cv.COLOR_BGR2GRAY)# 加载人脸分类识别器face_detector &＃61; cv.CascadeClassifier(r&＃39;../data/haarcascade_frontalface_default.xml&＃39;)faces &＃61; face_detector.detectMultiScale(frame_gray)for x, y, w, h in faces:# 头像标注# cv.rectangle(frame, (x, y), (x &＃43; w, y &＃43; h), color&＃61;(0, 255, 0), thickness&＃61;1)# 头像切割保存,分类型保存 frame_gray 为灰度图片 frame 为原始图片,选择灰度图片作为训练集cv.imwrite(set_path &＃43; str(uuid.uuid1()) &＃43; &＃39;.jpg&＃39;, frame_gray[y:y &＃43; h, x:x &＃43; w])if __name__ &＃61;&＃61; &＃39;__main__&＃39;:# 数据保存路径,不能有中文set_path &＃61; "C:/Users/Administrator/Desktop/data"label &＃61; 0 # 当前数据的标签stream_address &＃61; "../data/one.mp4" # 读取视频文件# stream_address &＃61; 0;//获取摄像头stream &＃61; cv.VideoCapture(stream_address) # 转化为流set_path &＃61; set_path &＃43; "/" &＃43; labelwhile True:status, frame &＃61; stream.read() # 读取帧图片数据if status:print("获取到图片数据,规格为:", frame.shape)else:print(&＃39;没有检查到图像数据&＃39;)breaksplit_image(frame, set_path) # 截取头部数据信息cv.destroyAllWindows()stream.release()

训练模型 train.py

# 训练模型
import numpy as np
from PIL import Image
import cv2 as cv
import os# 加载训练数据
def get_train_data(data_path, image_width, image_high):train_data_array &＃61; [] # 训练集,灰度图片label_array &＃61; [] # 标签 必须是整数,数值型从0开始包括0,后续识别可以通过标签索引获取数组字典,数组从0开始listdir &＃61; os.listdir(data_path)for label in listdir:label_path &＃61; data_path &＃43; "/" &＃43; labelfiles &＃61; os.listdir(label_path)for file_name in files:file_path &＃61; label_path &＃43; "/" &＃43; file_nameimg_numpy &＃61; np.array(Image.open(file_path).convert(&＃39;L&＃39;), &＃39;uint8&＃39;)img_format &＃61; cv.resize(img_numpy, (image_width, image_high), interpolation&＃61;cv.INTER_CUBIC) # 图片大小格式化train_data_array.append(img_format)label_array.append(int(label))return label_array, train_data_array# 训练模型
def training():print("人脸训练中 ...")data_path &＃61; "C:/Users/Administrator/Desktop/data" # 训练数据路径label_array, train_data_array &＃61; get_train_data(data_path, 128, 128) # 指定图片格式化大小# 创建LBPH识别器并开始训练&＃xff0c;当然也可以选择Eigen或者Fisher识别器recognizer &＃61; cv.face.LBPHFaceRecognizer_create()# 传入训练集与标签recognizer.train(train_data_array, np.array(label_array))# 保存训练信息recognizer.write(&＃39;../model/trainer.yml&＃39;)print("{0} 训练分类数为2,已完成".format(len(np.unique(label_array))))if __name__ &＃61;&＃61; &＃39;__main__&＃39;:training()

视频识别 distinguish.py

# 识别
import cv2 as cvdef distinguish():# 加载算法 创建LBPH识别器并开始识别&＃xff0c;当然也可以选择Eigen或者Fisher识别器recognizer &＃61; cv.face.LBPHFaceRecognizer_create()# 加载模型recognizer.read(&＃39;../model/trainer.yml&＃39;)face_detector &＃61; cv.CascadeClassifier(r"../data/haarcascade_frontalface_default.xml")font &＃61; cv.FONT_HERSHEY_SIMPLEX# 特征值对应的姓名映射names &＃61; [&＃39;one&＃39;, &＃39;two&＃39;]cap &＃61; cv.VideoCapture("../data/one.mp4") # 加载视频文件数据# cap &＃61; cv.VideoCapture(0) # 加载摄像头数据while True:status, frame &＃61; cap.read()if status:print("获取到图片数据,规格为:", frame.shape)else:print(&＃39;没有检查到图像数据&＃39;)break# 图片转灰度frame_gray &＃61; cv.cvtColor(frame, cv.COLOR_BGR2GRAY)faces &＃61; face_detector.detectMultiScale(frame_gray)for x, y, w, h in faces:cv.rectangle(frame, (x, y), (x &＃43; w, y &＃43; h), color&＃61;(0, 255, 0), thickness&＃61;1)# 识别feature_id, value &＃61; recognizer.predict(frame_gray[y:y &＃43; h, x:x &＃43; w])print("当前识别分类为", feature_id, "得分为", value)if value > 100:feature_id &＃61; "unknown"reliability &＃61; "100%"else:feature_id &＃61; names[feature_id] # 将识别的结果特征数值映射成姓名reliability &＃61; "{0}%".format(round(100 - value, 2))cv.putText(frame, str(feature_id), (x &＃43; 5, y - 5), font, 1, (0, 0, 255), 1)cv.putText(frame, str(reliability), (x &＃43; 5, y &＃43; h - 5), font, 1, (0, 255, 0), 1)cv.imshow("distinguish", frame)if ord(&＃39;q&＃39;) &＃61;&＃61; cv.waitKey(10):breakcv.destroyAllWindows()cap.release()if __name__ &＃61;&＃61; &＃39;__main__&＃39;:distinguish()


采集头像数据示例

训练完之后会在mian/model目录下生成训练文件

校验模型的准确度 check.py

# 模型校验
import cv2 as cv
import os# 数据结构
class CheckResult:accuracy &＃61; 0 # 正确数error &＃61; 0 # 正确数unknown &＃61; 0 # 未知数total &＃61; 0 # 总数accuracy_list &＃61; [] # 正确详细数据error_list &＃61; [] # 错误详细数据unknown_list &＃61; [] # 未知详细数据accuracy_rate &＃61; None # 正确率error_rate &＃61; None # 错误率unknown_rate &＃61; None # 未知率accuracy_reliability &＃61; 0 # 正确平均可信度error_reliability &＃61; 0 # 错误平均可信度def msg(self):return "正确识别数:{0},正确率:{1},错误识别数:{2},错误率:{3},未知识别数:{4},未知率:{5},正确可信度:{6},错误可信度:{7}".format(self.accuracy,self.accuracy_rate,self.error,self.error_rate, self.unknown,self.unknown_rate,self.accuracy_reliability,self.error_reliability)# 加载校验数据
def get_test_data(data_path):# 加载算法 创建LBPH识别器并开始校验&＃xff0c;当然也可以选择Eigen或者Fisher识别器recognizer &＃61; cv.face.LBPHFaceRecognizer_create() # 加载模型recognizer.read(&＃39;../model/trainer.yml&＃39;)face_detector &＃61; cv.CascadeClassifier(r"../data/haarcascade_frontalface_default.xml")test_data_array &＃61; [] # 校验集listdir &＃61; os.listdir(data_path)for label in listdir:label_path &＃61; data_path &＃43; "/" &＃43; labelfiles &＃61; os.listdir(label_path)for file_name in files:file_path &＃61; label_path &＃43; "/" &＃43; file_nameframe &＃61; cv.imread(file_path, 3)# 图片转灰度frame_gray &＃61; cv.cvtColor(frame, cv.COLOR_BGR2GRAY)faces &＃61; face_detector.detectMultiScale(frame_gray)for x, y, w, h in faces:cv.rectangle(frame, (x, y), (x &＃43; w, y &＃43; h), color&＃61;(0, 255, 0), thickness&＃61;1)# 识别feature_id, value &＃61; recognizer.predict(frame_gray[y:y &＃43; h, x:x &＃43; w])if value > 100:feature_id &＃61; "unknown"reliability &＃61; 1else:reliability &＃61; 100 - valuetest_data_array.append({"path": file_path, "feature": feature_id, "reliability": reliability, "label": int(label)})return test_data_array# 校验模型
def calc_result(test_data_array):result &＃61; CheckResult()for x in test_data_array:if x["feature"] &＃61;&＃61; x["label"]:result.accuracy &＃43;&＃61; 1result.accuracy_list.append(x)result.accuracy_reliability &＃43;&＃61; x["reliability"]elif x["feature"] &＃61;&＃61; &＃39;unknown&＃39;:result.unknown &＃43;&＃61; 1result.unknown_list.append(x)else:result.error &＃43;&＃61; 1result.error_list.append(x)result.error_reliability &＃43;&＃61; x["reliability"]result.total &＃61; len(test_data_array)if result.total &＃61;&＃61; 0:result.accuracy_rate &＃61; "0.00%"result.error_rate &＃61; "0.00%"result.unknown_rate &＃61; "0.00%"else:result.accuracy_rate &＃61; str(round(result.accuracy * 100 / result.total, 2)) &＃43; "%"result.error_rate &＃61; str(round(result.error * 100 / result.total, 2)) &＃43; "%"result.unknown_rate &＃61; str(round(result.unknown * 100 / result.total, 2)) &＃43; "%"result.accuracy_reliability &＃61; str(round(result.accuracy_reliability / result.total, 2)) &＃43; "%"result.error_reliability &＃61; str(round(result.error_reliability / result.total, 2)) &＃43; "%"return resultdef checking():print("模型校验中 ...")data_path &＃61; "C:/Users/Administrator/Desktop/data" # 校验数据路径test_data_array &＃61; get_test_data(data_path)result &＃61; calc_result(test_data_array)print(result.msg())# 校验模型准确度
if __name__ &＃61;&＃61; &＃39;__main__&＃39;:checking()


校验结果