前言一、opencv 人脸识别函数二、从素材准备到预测三、意外的发现四、结论
前言
学习opencv, 用网上找到的男女明星照骗试了下opencv的3种人脸识别机制, 似乎都不太行。
一、opencv 人脸识别函数
有3个识别器：
recognizer_eigenface = cv2.face.EigenFaceRecognizer_create()recognizer_fisher = cv2.face.FisherFaceRecognizer_create()recognizer_LBPH = cv2.face.LBPHFaceRecognizer_create()
使用方法是一样的，都要先准备好训练集来训练，训练后再预测注意安装标准版的opencv-python是不行的，要安装opencv-contrib-python才行。 二、从素材准备到预测 下载图片
男女明星各一组，每组2人，每人各9张。图片找小冰要的，度娘也一样。图片规整化
训练时要求图片尺寸统一，而下载来的图片大大小小，所以要统一到同一尺寸、用简单粗暴的方法，查出所有图片的最小尺寸，然后将所有照片调到此尺寸。不能变形，不足的方向补黑边。
def unifyImg(strSource): #read all images and get the smallest size x_min = 9999 y_min = 9999 img_Bingbing = list() img_file_list = os.listdir(strSource) for i in range(0,len(img_file_list)): strFileName = os.path.join(strSource,img_file_list[i]) if os.path.isfile(strFileName): img = cv2.imread(strFileName) if img is None: print('Failed when read file %s'%strFileName) continue # record size if img.shape[0] < y_min: y_min = img.shape[0] if img.shape[1] < x_min: x_min = img.shape[1] #store img and its name img_Bingbing.append([img,img_file_list[i]]) #Resize images frontFaceCascade = cv2.CascadeClassifier(r'.Libsite-packagescv2datahaarcascade_frontalface_default.xml') profileFaceCascade = cv2.CascadeClassifier(r'.Libsite-packagescv2datahaarcascade_profileface.xml') for i in range(0,len(img_Bingbing)): img = img_Bingbing[i][0] #check if face is identified in the image frontFaces = frontFaceCascade.detectMultiScale(img) profileFaces = profileFaceCascade.detectMultiScale(img) if len(frontFaces) == 0: if len(profileFaces) == 0: print("no face detected on img: %s, deleted."%img_Bingbing[i][1]) continue else: x = profileFaces[0][0] y = profileFaces[0][1] w = profileFaces[0][2] h = profileFaces[0][3] else: x = frontFaces[0][0] y = frontFaces[0][1] w = frontFaces[0][2] h = frontFaces[0][3] img_new = img[y:y+h,x:x+w] cv2.imwrite(strSource + "\Face_only\" + img_Bingbing[i][1],img_new) # 98% make sure the output image size is smaller than the min_x or min_y rate_x = x_min / img.shape[1] * 0.98 rate_y = y_min / img.shape[0] * 0.98 rate = min(rate_x, rate_y) img_new = cv2.resize(img, None, fx = rate, fy = rate) if len(img_new.shape) == 2: # grey img_template = np.zeros((y_min, x_min),np.uint8) else: img_template = np.zeros((y_min, x_min, 3),np.uint8) h = img_new.shape[0] w = img_new.shape[1] img_template[0:h,0:w] = img_new cv2.imwrite(strSource + "\Unified\" + img_Bingbing[i][1],img_template)strSource = r'C:WaleyPersonalLearningPythonImageTryFaceIdentify'unifyImg(strSource)
挑一部分统一好尺寸的图片整合进一个列表，用来训练src = list()lables = list()# LiYiFengfor i in range(0,6): img = cv2.imread(strSource + "\unified\LYF_" + str(i+1) + ".jpg",cv2.IMREAD_GRAYSCALE) if img is None: print("Fail to read images") sys.exit(0) else: src.append(img) lables.append(0)#YiYangQianXifor i in range(0,6): img = cv2.imread(strSource + "\unified\YYQX_" + str(i+1) + ".jpg",cv2.IMREAD_GRAYSCALE) if img is None: print("Fail to read images") sys.exit(0) else: src.append(img) lables.append(1) names = {"0":"LiYiFeng","1":"YiYangQianXi"}#train recognizerrecognizer_eigenface.train(src,np.array(lables))recognizer_fisher.train(src, np.array(lables))recognizer_LBPH.train(src, np.array(lables))
识别for i in range(7,10): img = cv2.imread(strSource + "\unified\LYF_" + str(i) + '.jpg',cv2.IMREAD_GRAYSCALE) print("LYF %d : "%i) lable, confidence = recognizer_eigenface.predict(img) print("predicted by Eigenface: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_fisher.predict(img) print("predicted by Fisher: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_LBPH.predict(img) print("predicted by LBPH: %s, confidence is: %d"%(names[str(lable)],confidence))print("nn")for i in range(7,10): img = cv2.imread(strSource + "\unified\YYQX_" + str(i) + '.jpg',cv2.IMREAD_GRAYSCALE) print("YYQX %d : "%i) lable, confidence = recognizer_eigenface.predict(img) print("predicted by Eigenface: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_fisher.predict(img) print("predicted by Fisher: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_LBPH.predict(img) print("predicted by LBPH: %s, confidence is: %d"%(names[str(lable)],confidence))
以上可以合并为一个函数，方便换人试验:def trainAndPredict(strSource, strName1, strName2): # prepare training dataset src = list() lables = list() for i in range(0,6): img = cv2.imread(strSource + "\unified\" + strName1 + '_' + str(i+1) + ".jpg",cv2.IMREAD_GRAYSCALE) if img is None: print("Fail to read images") sys.exit(0) else: src.append(img) lables.append(0) for i in range(0,6): img = cv2.imread(strSource + "\unified\" + strName2 + '_' + str(i+1) + ".jpg",cv2.IMREAD_GRAYSCALE) if img is None: print("Fail to read images") sys.exit(0) else: src.append(img) lables.append(1) names = {"0":strName1,"1":strName2} #set up creator recognizer_eigenface = cv2.face.EigenFaceRecognizer_create() recognizer_fisher = cv2.face.FisherFaceRecognizer_create() recognizer_LBPH = cv2.face.LBPHFaceRecognizer_create() #train recognizer recognizer_eigenface.train(src,np.array(lables)) recognizer_fisher.train(src, np.array(lables)) recognizer_LBPH.train(src, np.array(lables)) for i in range(7,10): img = cv2.imread(strSource + "\unified\" + strName1 + '_' + str(i) + '.jpg',cv2.IMREAD_GRAYSCALE) print("%s %d : "%(strName1,i)) lable, confidence = recognizer_eigenface.predict(img) print("predicted by Eigenface: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_fisher.predict(img) print("predicted by Fisher: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_LBPH.predict(img) print("predicted by LBPH: %s, confidence is: %d"%(names[str(lable)],confidence)) print("nn") for i in range(7,10): img = cv2.imread(strSource + "\unified\" + strName2 + '_' + str(i) + '.jpg',cv2.IMREAD_GRAYSCALE) print("%s %d : "%(strName2,i)) lable, confidence = recognizer_eigenface.predict(img) print("predicted by Eigenface: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_fisher.predict(img) print("predicted by Fisher: %s, confidence is: %d"%(names[str(lable)],confidence)) lable, confidence = recognizer_LBPH.predict(img) print("predicted by LBPH: %s, confidence is: %d"%(names[str(lable)],confidence))#strSource = r'C:WaleyPersonalLearningPythonImageTryFaceIdentify'#strName1 = 'LYF'#strName2 = 'YYQX'strSource = r'C:WaleyPersonalLearningPythonImageFaceIdentify'strName1 = 'LiBingbing'strName2 = 'YangMi'trainAndPredict(strSource, strName1, strName2)
结果让人意外对于两位男明星，简直没有任何识别率。
这两位明星都不太熟悉，会不会是网站也是AI标的，标错了呢？哈哈。
换了两位女明星试验，效果似乎还行：
发现这张图片opencv 检测不出来脸部，正面侧面都不行。等继续研究为什么。
发现这张图片opencv识别出来的脸部是错误的
识别出来的脸是肘部，
原因待进一步研究。
四、结论将opencv 的人脸识别用于生产环境估计是不行的。下一步准备用同样的这些图片训练下TensorFlow试试