基于深度学习的目标检测/人脸检测/人脸识别：MTCNN、基于度量学习的方法，基于边界分类的方法

向右看齐 2023-07-10 12:57 146阅读 0赞

# ![20191009191333910.png][][日萌社][Link 1] #

[人工智能AI：Keras PyTorch MXNet TensorFlow PaddlePaddle 深度学习实战（不定时更新）][AI_Keras PyTorch MXNet TensorFlow PaddlePaddle]

--------------------

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70][] ![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 1][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 2][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 3][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 4][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 5][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 6][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 7][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 8][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 9][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 10][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 11][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 12][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 13][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 14][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 15][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 16][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 17][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 18][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 19][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 20][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 21][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 22][]

mtcnn_model.py
    
    from keras.layers import Conv2D, Input, MaxPool2D, Flatten, Dense, Permute
    from keras.layers.advanced_activations import PReLU
    from keras.models import Model
    
    
    def create_Kao_Onet( weight_path = 'model48.h5'):
        input = Input(shape = [48,48,3])
        x = Conv2D(32, (3, 3), strides=1, padding='valid', name='conv1')(input)
        x = PReLU(shared_axes=[1,2],name='prelu1')(x)
        x = MaxPool2D(pool_size=3, strides=2, padding='same')(x)
        x = Conv2D(64, (3, 3), strides=1, padding='valid', name='conv2')(x)
        x = PReLU(shared_axes=[1,2],name='prelu2')(x)
        x = MaxPool2D(pool_size=3, strides=2)(x)
        x = Conv2D(64, (3, 3), strides=1, padding='valid', name='conv3')(x)
        x = PReLU(shared_axes=[1,2],name='prelu3')(x)
        x = MaxPool2D(pool_size=2)(x)
        x = Conv2D(128, (2, 2), strides=1, padding='valid', name='conv4')(x)
        x = PReLU(shared_axes=[1,2],name='prelu4')(x)
        x = Permute((3,2,1))(x)
        x = Flatten()(x)
        x = Dense(256, name='conv5') (x)
        x = PReLU(name='prelu5')(x)
    
        classifier = Dense(2, activation='softmax',name='conv6-1')(x)
        bbox_regress = Dense(4,name='conv6-2')(x)
        landmark_regress = Dense(10,name='conv6-3')(x)
        model = Model([input], [classifier, bbox_regress, landmark_regress])
        model.load_weights(weight_path, by_name=True)
    
        return model
    
    
    def create_Kao_Rnet (weight_path = 'model24.h5'):
        input = Input(shape=[24, 24, 3])  # change this shape to [None,None,3] to enable arbitraty shape input
        x = Conv2D(28, (3, 3), strides=1, padding='valid', name='conv1')(input)
        x = PReLU(shared_axes=[1, 2], name='prelu1')(x)
        x = MaxPool2D(pool_size=3,strides=2, padding='same')(x)
    
        x = Conv2D(48, (3, 3), strides=1, padding='valid', name='conv2')(x)
        x = PReLU(shared_axes=[1, 2], name='prelu2')(x)
        x = MaxPool2D(pool_size=3, strides=2)(x)
    
        x = Conv2D(64, (2, 2), strides=1, padding='valid', name='conv3')(x)
        x = PReLU(shared_axes=[1, 2], name='prelu3')(x)
        x = Permute((3, 2, 1))(x)
        x = Flatten()(x)
        x = Dense(128, name='conv4')(x)
        x = PReLU( name='prelu4')(x)
        classifier = Dense(2, activation='softmax', name='conv5-1')(x)
        bbox_regress = Dense(4, name='conv5-2')(x)
        model = Model([input], [classifier, bbox_regress])
        model.load_weights(weight_path, by_name=True)
        return model
    
    
    def create_Kao_Pnet( weight_path = 'model12old.h5'):
        input = Input(shape=[None, None, 3])
        x = Conv2D(10, (3, 3), strides=1, padding='valid', name='conv1')(input)
        x = PReLU(shared_axes=[1,2],name='PReLU1')(x)
        x = MaxPool2D(pool_size=2)(x)
        x = Conv2D(16, (3, 3), strides=1, padding='valid', name='conv2')(x)
        x = PReLU(shared_axes=[1,2],name='PReLU2')(x)
        x = Conv2D(32, (3, 3), strides=1, padding='valid', name='conv3')(x)
        x = PReLU(shared_axes=[1,2],name='PReLU3')(x)
        classifier = Conv2D(2, (1, 1), activation='softmax', name='conv4-1')(x)
        bbox_regress = Conv2D(4, (1, 1), name='conv4-2')(x)
        model = Model([input], [classifier, bbox_regress])
        model.load_weights(weight_path, by_name=True)
        return model

Run_model_caffe_weight.py
    
    import sys
    import mtcnn.tools_matrix as tools
    import cv2
    import numpy as np
    import matplotlib.pyplot as plt
    import time
    from mtcnn.mtcnn_model import create_Kao_Onet, create_Kao_Rnet, create_Kao_Pnet
    
    
    Pnet = create_Kao_Pnet(r'12net.h5')
    Rnet = create_Kao_Rnet(r'24net.h5')
    Onet = create_Kao_Onet(r'48net.h5')  # will not work. caffe and TF incompatible
    
    
    def detectFace(img, threshold):
    
        caffe_img = (img.copy() - 127.5) / 127.5
        origin_h, origin_w, ch = caffe_img.shape
        scales = tools.calculateScales(img)
        out = []
        t0 = time.time()
        # del scales[:4]
    
        for scale in scales:
            hs = int(origin_h * scale)
            ws = int(origin_w * scale)
            scale_img = cv2.resize(caffe_img, (ws, hs))
            input = scale_img.reshape(1, *scale_img.shape)
            ouput = Pnet.predict(input)  # .transpose(0,2,1,3) should add, but seems after process is wrong then.
            out.append(ouput)
        image_num = len(scales)
        rectangles = []
        for i in range(image_num):
            cls_prob = out[i][0][0][:, :,
                       1]  # i = #scale, first 0 select cls score, second 0 = batchnum, alway=0. 1 one hot repr
            roi = out[i][1][0]
            out_h, out_w = cls_prob.shape
            out_side = max(out_h, out_w)
            # print('calculating img scale #:', i)
            cls_prob = np.swapaxes(cls_prob, 0, 1)
            roi = np.swapaxes(roi, 0, 2)
            rectangle = tools.detect_face_12net(cls_prob, roi, out_side, 1 / scales[i], origin_w, origin_h, threshold[0])
            rectangles.extend(rectangle)
        rectangles = tools.NMS(rectangles, 0.7, 'iou')
    
        t1 = time.time()
        print ('time for 12 net is: ', t1-t0)
    
        if len(rectangles) == 0:
            return rectangles
    
        crop_number = 0
        out = []
        predict_24_batch = []
        for rectangle in rectangles:
            crop_img = caffe_img[int(rectangle[1]):int(rectangle[3]), int(rectangle[0]):int(rectangle[2])]
            scale_img = cv2.resize(crop_img, (24, 24))
            predict_24_batch.append(scale_img)
            crop_number += 1
    
        predict_24_batch = np.array(predict_24_batch)
    
        out = Rnet.predict(predict_24_batch)
    
        cls_prob = out[0]  # first 0 is to select cls, second batch number, always =0
        cls_prob = np.array(cls_prob)  # convert to numpy
        roi_prob = out[1]  # first 0 is to select roi, second batch number, always =0
        roi_prob = np.array(roi_prob)
        rectangles = tools.filter_face_24net(cls_prob, roi_prob, rectangles, origin_w, origin_h, threshold[1])
        t2 = time.time()
        print ('time for 24 net is: ', t2-t1)
    
    
        if len(rectangles) == 0:
            return rectangles
    
    
        crop_number = 0
        predict_batch = []
        for rectangle in rectangles:
            # print('calculating net 48 crop_number:', crop_number)
            crop_img = caffe_img[int(rectangle[1]):int(rectangle[3]), int(rectangle[0]):int(rectangle[2])]
            scale_img = cv2.resize(crop_img, (48, 48))
            predict_batch.append(scale_img)
            crop_number += 1
    
        predict_batch = np.array(predict_batch)
    
        output = Onet.predict(predict_batch)
        cls_prob = output[0]
        roi_prob = output[1]
        pts_prob = output[2]  # index
        # rectangles = tools.filter_face_48net_newdef(cls_prob, roi_prob, pts_prob, rectangles, origin_w, origin_h,
        #                                             threshold[2])
        rectangles = tools.filter_face_48net(cls_prob, roi_prob, pts_prob, rectangles, origin_w, origin_h, threshold[2])
        t3 = time.time()
        print ('time for 48 net is: ', t3-t2)
    
        return rectangles
    
    
    threshold = [0.6,0.6,0.7]
    # video_path = 'WalmartArguments_p1.mkv'
    # cap = cv2.VideoCapture(video_path)
    
    while (True):
        # ret, img = cap.read()
        img = cv2.imread('manual_testing.jpeg')
    
        rectangles = detectFace(img, threshold)
        print(rectangles)
        draw = img.copy()
    
        for rectangle in rectangles:
            if rectangle is not None:
                W = -int(rectangle[0]) + int(rectangle[2])
                H = -int(rectangle[1]) + int(rectangle[3])
                paddingH = 0.01 * W
                paddingW = 0.02 * H
                crop_img = img[int(rectangle[1]+paddingH):int(rectangle[3]-paddingH), int(rectangle[0]-paddingW):int(rectangle[2]+paddingW)]
                crop_img = cv2.cvtColor(crop_img, cv2.COLOR_RGB2GRAY)
                if crop_img is None:
                    continue
                if crop_img.shape[0] < 0 or crop_img.shape[1] < 0:
                    continue
                cv2.rectangle(draw, (int(rectangle[0]), int(rectangle[1])), (int(rectangle[2]), int(rectangle[3])), (255, 0, 0), 1)
    
                for i in range(5, 15, 2):
                    cv2.circle(draw, (int(rectangle[i + 0]), int(rectangle[i + 1])), 2, (0, 255, 0))
    
        cv2.imwrite('0001.jpg', draw)
        break

tools_matrix.py
    
    import sys
    from operator import itemgetter
    import numpy as np
    import cv2
    
    '''
    Function:
        change rectangles into squares (matrix version)
    Input:
        rectangles: rectangles[i][0:3] is the position, rectangles[i][4] is score
    Output:
        squares: same as input
    '''
    def rect2square(rectangles):
        w = rectangles[:,2] - rectangles[:,0]
        h = rectangles[:,3] - rectangles[:,1]
        l = np.maximum(w,h).T
        rectangles[:,0] = rectangles[:,0] + w*0.5 - l*0.5
        rectangles[:,1] = rectangles[:,1] + h*0.5 - l*0.5 
        rectangles[:,2:4] = rectangles[:,0:2] + np.repeat([l], 2, axis = 0).T 
        return rectangles
    '''
    Function:
        apply NMS(non-maximum suppression) on ROIs in same scale(matrix version)
    Input:
        rectangles: rectangles[i][0:3] is the position, rectangles[i][4] is score
    Output:
        rectangles: same as input
    '''
    def NMS(rectangles,threshold,type):
        if len(rectangles)==0:
            return rectangles
        boxes = np.array(rectangles)
        x1 = boxes[:,0]
        y1 = boxes[:,1]
        x2 = boxes[:,2]
        y2 = boxes[:,3]
        s  = boxes[:,4]
        area = np.multiply(x2-x1+1, y2-y1+1)
        I = np.array(s.argsort())
        pick = []
        while len(I)>0:
            xx1 = np.maximum(x1[I[-1]], x1[I[0:-1]]) #I[-1] have hightest prob score, I[0:-1]->others
            yy1 = np.maximum(y1[I[-1]], y1[I[0:-1]])
            xx2 = np.minimum(x2[I[-1]], x2[I[0:-1]])
            yy2 = np.minimum(y2[I[-1]], y2[I[0:-1]])
            w = np.maximum(0.0, xx2 - xx1 + 1)
            h = np.maximum(0.0, yy2 - yy1 + 1)
            inter = w * h
            if type == 'iom':
                o = inter / np.minimum(area[I[-1]], area[I[0:-1]])
            else:
                o = inter / (area[I[-1]] + area[I[0:-1]] - inter)
            pick.append(I[-1])
            I = I[np.where(o<=threshold)[0]]
        result_rectangle = boxes[pick].tolist()
        return result_rectangle
    '''
    Function:
        Detect face position and calibrate bounding box on 12net feature map(matrix version)
    Input:
        cls_prob : softmax feature map for face classify
        roi      : feature map for regression
        out_side : feature map's largest size
        scale    : current input image scale in multi-scales
        width    : image's origin width
        height   : image's origin height
        threshold: 0.6 can have 99% recall rate
    '''
    def detect_face_12net(cls_prob,roi,out_side,scale,width,height,threshold):
        in_side = 2*out_side+11
        stride = 0
        if out_side != 1:
            stride = float(in_side-12)/(out_side-1)
        (x,y) = np.where(cls_prob>=threshold)
        boundingbox = np.array([x,y]).T
        bb1 = np.fix((stride * (boundingbox) + 0 ) * scale)
        bb2 = np.fix((stride * (boundingbox) + 11) * scale)
        boundingbox = np.concatenate((bb1,bb2),axis = 1)
        dx1 = roi[0][x,y]
        dx2 = roi[1][x,y]
        dx3 = roi[2][x,y]
        dx4 = roi[3][x,y]
        score = np.array([cls_prob[x,y]]).T
        offset = np.array([dx1,dx2,dx3,dx4]).T
        boundingbox = boundingbox + offset*12.0*scale
        rectangles = np.concatenate((boundingbox,score),axis=1)
        rectangles = rect2square(rectangles)
        pick = []
        for i in range(len(rectangles)):
            x1 = int(max(0     ,rectangles[i][0]))
            y1 = int(max(0     ,rectangles[i][1]))
            x2 = int(min(width ,rectangles[i][2]))
            y2 = int(min(height,rectangles[i][3]))
            sc = rectangles[i][4]
            if x2>x1 and y2>y1:
                pick.append([x1,y1,x2,y2,sc])
        return NMS(pick,0.3,'iou')
    '''
    Function:
        Filter face position and calibrate bounding box on 12net's output
    Input:
        cls_prob  : softmax feature map for face classify
        roi_prob  : feature map for regression
        rectangles: 12net's predict
        width     : image's origin width
        height    : image's origin height
        threshold : 0.6 can have 97% recall rate
    Output:
        rectangles: possible face positions
    '''
    def filter_face_24net(cls_prob,roi,rectangles,width,height,threshold):
        prob = cls_prob[:,1]
        pick = np.where(prob>=threshold)
        rectangles = np.array(rectangles)
        x1  = rectangles[pick,0]
        y1  = rectangles[pick,1]
        x2  = rectangles[pick,2]
        y2  = rectangles[pick,3]
        sc  = np.array([prob[pick]]).T
        dx1 = roi[pick,0]
        dx2 = roi[pick,1]
        dx3 = roi[pick,2]
        dx4 = roi[pick,3]
        w   = x2-x1
        h   = y2-y1
        x1  = np.array([(x1+dx1*w)[0]]).T
        y1  = np.array([(y1+dx2*h)[0]]).T
        x2  = np.array([(x2+dx3*w)[0]]).T
        y2  = np.array([(y2+dx4*h)[0]]).T
        rectangles = np.concatenate((x1,y1,x2,y2,sc),axis=1)
        rectangles = rect2square(rectangles)
        pick = []
        for i in range(len(rectangles)):
            x1 = int(max(0     ,rectangles[i][0]))
            y1 = int(max(0     ,rectangles[i][1]))
            x2 = int(min(width ,rectangles[i][2]))
            y2 = int(min(height,rectangles[i][3]))
            sc = rectangles[i][4]
            if x2>x1 and y2>y1:
                pick.append([x1,y1,x2,y2,sc])
        return NMS(pick,0.3,'iou')
    '''
    Function:
        Filter face position and calibrate bounding box on 12net's output
    Input:
        cls_prob  : cls_prob[1] is face possibility
        roi       : roi offset
        pts       : 5 landmark
        rectangles: 12net's predict, rectangles[i][0:3] is the position, rectangles[i][4] is score
        width     : image's origin width
        height    : image's origin height
        threshold : 0.7 can have 94% recall rate on CelebA-database
    Output:
        rectangles: face positions and landmarks
    '''
    def filter_face_48net(cls_prob,roi,pts,rectangles,width,height,threshold):
        prob = cls_prob[:,1]
        pick = np.where(prob>=threshold)
        rectangles = np.array(rectangles)
        x1  = rectangles[pick,0]
        y1  = rectangles[pick,1]
        x2  = rectangles[pick,2]
        y2  = rectangles[pick,3]
        sc  = np.array([prob[pick]]).T
        dx1 = roi[pick,0]
        dx2 = roi[pick,1]
        dx3 = roi[pick,2]
        dx4 = roi[pick,3]
        w   = x2-x1
        h   = y2-y1
        pts0= np.array([(w*pts[pick,0]+x1)[0]]).T
        pts1= np.array([(h*pts[pick,5]+y1)[0]]).T
        pts2= np.array([(w*pts[pick,1]+x1)[0]]).T
        pts3= np.array([(h*pts[pick,6]+y1)[0]]).T
        pts4= np.array([(w*pts[pick,2]+x1)[0]]).T
        pts5= np.array([(h*pts[pick,7]+y1)[0]]).T
        pts6= np.array([(w*pts[pick,3]+x1)[0]]).T
        pts7= np.array([(h*pts[pick,8]+y1)[0]]).T
        pts8= np.array([(w*pts[pick,4]+x1)[0]]).T
        pts9= np.array([(h*pts[pick,9]+y1)[0]]).T
        # pts0 = np.array([(w * pts[pick, 0] + x1)[0]]).T
        # pts1 = np.array([(h * pts[pick, 1] + y1)[0]]).T
        # pts2 = np.array([(w * pts[pick, 2] + x1)[0]]).T
        # pts3 = np.array([(h * pts[pick, 3] + y1)[0]]).T
        # pts4 = np.array([(w * pts[pick, 4] + x1)[0]]).T
        # pts5 = np.array([(h * pts[pick, 5] + y1)[0]]).T
        # pts6 = np.array([(w * pts[pick, 6] + x1)[0]]).T
        # pts7 = np.array([(h * pts[pick, 7] + y1)[0]]).T
        # pts8 = np.array([(w * pts[pick, 8] + x1)[0]]).T
        # pts9 = np.array([(h * pts[pick, 9] + y1)[0]]).T
        x1  = np.array([(x1+dx1*w)[0]]).T
        y1  = np.array([(y1+dx2*h)[0]]).T
        x2  = np.array([(x2+dx3*w)[0]]).T
        y2  = np.array([(y2+dx4*h)[0]]).T
        rectangles=np.concatenate((x1,y1,x2,y2,sc,pts0,pts1,pts2,pts3,pts4,pts5,pts6,pts7,pts8,pts9),axis=1)
        pick = []
        for i in range(len(rectangles)):
            x1 = int(max(0     ,rectangles[i][0]))
            y1 = int(max(0     ,rectangles[i][1]))
            x2 = int(min(width ,rectangles[i][2]))
            y2 = int(min(height,rectangles[i][3]))
            if x2>x1 and y2>y1:
                pick.append([x1,y1,x2,y2,rectangles[i][4],
                     rectangles[i][5],rectangles[i][6],rectangles[i][7],rectangles[i][8],rectangles[i][9],rectangles[i][10],rectangles[i][11],rectangles[i][12],rectangles[i][13],rectangles[i][14]])
        return NMS(pick,0.3,'iom')
    '''
    Function:
        calculate multi-scale and limit the maxinum side to 1000 
    Input: 
        img: original image
    Output:
        pr_scale: limit the maxinum side to 1000, < 1.0
        scales  : Multi-scale
    '''
    def calculateScales(img):
        caffe_img = img.copy()
        pr_scale = 1.0
        h,w,ch = caffe_img.shape
        if min(w,h)>500:
            pr_scale = 500.0/min(h,w)
            w = int(w*pr_scale)
            h = int(h*pr_scale)
        elif max(w,h)<500:
            pr_scale = 500.0/max(h,w)
            w = int(w*pr_scale)
            h = int(h*pr_scale)
    
        #multi-scale
        scales = []
        factor = 0.709
        factor_count = 0
        minl = min(h,w)
        while minl >= 12:
            scales.append(pr_scale*pow(factor, factor_count))
            minl *= factor
            factor_count += 1
        return scales
    
    # '''
    # Function:
    #     calculate switch definition of landmark point to new def
    # Input:
    #     pts: old definition pts
    # Output:
    #     pts_new: new def pts
    # '''
    # def pts_def_rectify(pts):
    #     pts_new = np.zeros_like(pts)
    #     pts_new[:, 0]= pts[:,0]
    #     pts_new[:, 1]= pts[:,5]
    #     pts_new[:, 2]= pts[:,1]
    #     pts_new[:, 3]= pts[:,6]
    #     pts_new[:, 4]= pts[:,2]
    #     pts_new[:, 5]= pts[:,7]
    #     pts_new[:, 6]= pts[:,3]
    #     pts_new[:, 7]= pts[:,8]
    #     pts_new[:, 8]= pts[:,4]
    #     pts_new[:, 9]= pts[:,9]
    #     return pts_new
    
    '''
    Function:
        calculate   landmark point , new def
    Input:
        cls_prob  : cls_prob[1] is face possibility
        roi       : roi offset
        pts       : 5 landmark
        rectangles: 12net's predict, rectangles[i][0:3] is the position, rectangles[i][4] is score
        width     : image's origin width
        height    : image's origin height
        threshold : 0.7 can have 94% recall rate on CelebA-database
    Output:
        rectangles: face positions and landmarks
    '''
    
    
    def filter_face_48net_newdef(cls_prob,roi,pts,rectangles,width,height,threshold):
        prob = cls_prob[:,1]
        pick = np.where(prob>=threshold)
        rectangles = np.array(rectangles)
        x1  = rectangles[pick,0]
        y1  = rectangles[pick,1]
        x2  = rectangles[pick,2]
        y2  = rectangles[pick,3]
        sc  = np.array([prob[pick]]).T
        dx1 = roi[pick,0]
        dx2 = roi[pick,1]
        dx3 = roi[pick,2]
        dx4 = roi[pick,3]
        w   = x2-x1
        h   = y2-y1
        pts0= np.array([(w*pts[pick,0]+x1)[0]]).T
        pts1= np.array([(h*pts[pick,1]+y1)[0]]).T
        pts2= np.array([(w*pts[pick,2]+x1)[0]]).T
        pts3= np.array([(h*pts[pick,3]+y1)[0]]).T
        pts4= np.array([(w*pts[pick,4]+x1)[0]]).T
        pts5= np.array([(h*pts[pick,5]+y1)[0]]).T
        pts6= np.array([(w*pts[pick,6]+x1)[0]]).T
        pts7= np.array([(h*pts[pick,7]+y1)[0]]).T
        pts8= np.array([(w*pts[pick,8]+x1)[0]]).T
        pts9= np.array([(h*pts[pick,9]+y1)[0]]).T
        x1  = np.array([(x1+dx1*w)[0]]).T
        y1  = np.array([(y1+dx2*h)[0]]).T
        x2  = np.array([(x2+dx3*w)[0]]).T
        y2  = np.array([(y2+dx4*h)[0]]).T
        rectangles=np.concatenate((x1,y1,x2,y2,sc,pts0,pts1,pts2,pts3,pts4,pts5,pts6,pts7,pts8,pts9),axis=1)
        # print (pts0,pts1,pts2,pts3,pts4,pts5,pts6,pts7,pts8,pts9)
        pick = []
        for i in range(len(rectangles)):
            x1 = int(max(0     ,rectangles[i][0]))
            y1 = int(max(0     ,rectangles[i][1]))
            x2 = int(min(width ,rectangles[i][2]))
            y2 = int(min(height,rectangles[i][3]))
            if x2>x1 and y2>y1:
                pick.append([x1,y1,x2,y2,rectangles[i][4],
                     rectangles[i][5],rectangles[i][6],rectangles[i][7],rectangles[i][8],rectangles[i][9],rectangles[i][10],rectangles[i][11],rectangles[i][12],rectangles[i][13],rectangles[i][14]])
        return NMS(pick,0.3,'idsom')
    '''
    Function:
        calculate mean value of img_list for double checck img quality
    Input:
        img_nparray: numpy array of input
    Output:
        img_nparray: numpy array of img mean value
    '''
    
    
    def imglist_meanvalue(img_nparray):
        img_mean_array = np.mean(img_nparray ,axis=(1,2,3))
        return np.array(img_mean_array)

![20200303075257204.png][]

[20191009191333910.png]: https://img-blog.csdnimg.cn/20191009191333910.png
[Link 1]: http://www.rimengshe.com/
[AI_Keras PyTorch MXNet TensorFlow PaddlePaddle]: https://blog.csdn.net/zimiao552147572/article/details/88867161
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200302190742386.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20200302190829409.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20200302190839645.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20200302190844940.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20200302190853209.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 5]: https://img-blog.csdnimg.cn/20200302190857911.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 6]: https://img-blog.csdnimg.cn/20200302190902415.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 7]: https://img-blog.csdnimg.cn/20200302190906928.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 8]: https://img-blog.csdnimg.cn/20200302190911759.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 9]: https://img-blog.csdnimg.cn/20200302190916531.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 10]: https://img-blog.csdnimg.cn/20200302190921305.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 11]: https://img-blog.csdnimg.cn/20200302190925250.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 12]: https://img-blog.csdnimg.cn/20200302190930256.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 13]: https://img-blog.csdnimg.cn/20200302190934426.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 14]: https://img-blog.csdnimg.cn/20200302190939281.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 15]: https://img-blog.csdnimg.cn/20200302190943816.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 16]: https://img-blog.csdnimg.cn/20200302190948683.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 17]: https://img-blog.csdnimg.cn/20200302190954471.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 18]: https://img-blog.csdnimg.cn/20200302191002844.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 19]: https://img-blog.csdnimg.cn/20200302191008721.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 20]: https://img-blog.csdnimg.cn/20200302191014863.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 21]: https://img-blog.csdnimg.cn/20200302191019332.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg_size_16_color_FFFFFF_t_70 22]: https://img-blog.csdnimg.cn/20200302191023250.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3ppbWlhbzU1MjE0NzU3Mg==,size_16,color_FFFFFF,t_70
[20200303075257204.png]: https://img-blog.csdnimg.cn/20200303075257204.png