02-人脸识别-基于MTCNN，框选人脸区域-detect_face

（本系列随笔持续更新）

这部分代码是基于参考中的链接，修改后适用于TensorFlow1.6.0版本的代码。由于TensorFlow的频繁更新，所以不一定支持后续新或者就版本，特此说明。

程序的最初版，来自“山人7” [参考1，参考2]，但是在新的TensorFlow下面不能直接运行。

修改后版本，来自ShyBigBoy，[参考3，参考4]，可以在TensorFlow1.6.0上运行。

 

代码：detect_face.py

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780

""" Tensorflow implementation of the face detection / alignment algorithm found at https://github.com/kpzhang93/MTCNN_face_detection_alignment """ # MIT License # # Copyright (c) 2016 David Sandberg # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE.   from __future__ import absolute_import from __future__ import division from __future__ import print_function from six import string_types, iteritems   import numpy as np import tensorflow as tf #from math import floor import cv2 import os   def layer(op):     """Decorator for composable network layers."""       def layer_decorated(self, *args, **kwargs):         # Automatically set a name if not provided.         name = kwargs.setdefault('name', self.get_unique_name(op.__name__))         # Figure out the layer inputs.         if len(self.terminals) == 0:             raise RuntimeError('No input variables found for layer %s.' % name)         elif len(self.terminals) == 1:             layer_input = self.terminals[0]         else:             layer_input = list(self.terminals)         # Perform the operation and get the output.         layer_output = op(self, layer_input, *args, **kwargs)         # Add to layer LUT.         self.layers[name] = layer_output         # This output is now the input for the next layer.         self.feed(layer_output)         # Return self for chained calls.         return self       return layer_decorated   class Network(object):       def __init__(self, inputs, trainable=True):         # The input nodes for this network         self.inputs = inputs         # The current list of terminal nodes         self.terminals = []         # Mapping from layer names to layers         self.layers = dict(inputs)         # If true, the resulting variables are set as trainable         self.trainable = trainable           self.setup()       def setup(self):         """Construct the network. """         raise NotImplementedError('Must be implemented by the subclass.')       def load(self, data_path, session, ignore_missing=False):         """Load network weights.         data_path: The path to the numpy-serialized network weights         session: The current TensorFlow session         ignore_missing: If true, serialized weights for missing layers are ignored.         """         data_dict = np.load(data_path, encoding='latin1').item() #pylint: disable=no-member           for op_name in data_dict:             with tf.variable_scope(op_name, reuse=True):                 for param_name, data in iteritems(data_dict[op_name]):                     try:                         var = tf.get_variable(param_name)                         session.run(var.assign(data))                     except ValueError:                         if not ignore_missing:                             raise       def feed(self, *args):         """Set the input(s) for the next operation by replacing the terminal nodes.         The arguments can be either layer names or the actual layers.         """         assert len(args) != 0         self.terminals = []         for fed_layer in args:             if isinstance(fed_layer, string_types):                 try:                     fed_layer = self.layers[fed_layer]                 except KeyError:                     raise KeyError('Unknown layer name fed: %s' % fed_layer)             self.terminals.append(fed_layer)         return self       def get_output(self):         """Returns the current network output."""         return self.terminals[-1]       def get_unique_name(self, prefix):         """Returns an index-suffixed unique name for the given prefix.         This is used for auto-generating layer names based on the type-prefix.         """         ident = sum(t.startswith(prefix) for t, _ in self.layers.items()) + 1         return '%s_%d' % (prefix, ident)       def make_var(self, name, shape):         """Creates a new TensorFlow variable."""         return tf.get_variable(name, shape, trainable=self.trainable)       def validate_padding(self, padding):         """Verifies that the padding is one of the supported ones."""         assert padding in ('SAME', 'VALID')       @layer     def conv(self,              inp,              k_h,              k_w,              c_o,              s_h,              s_w,              name,              relu=True,              padding='SAME',              group=1,              biased=True):         # Verify that the padding is acceptable         self.validate_padding(padding)         # Get the number of channels in the input         c_i = int(inp.get_shape()[-1])         # Verify that the grouping parameter is valid         assert c_i % group == 0         assert c_o % group == 0         # Convolution for a given input and kernel         convolve = lambda i, k: tf.nn.conv2d(i, k, [1, s_h, s_w, 1], padding=padding)         with tf.variable_scope(name) as scope:             kernel = self.make_var('weights', shape=[k_h, k_w, c_i // group, c_o])             # This is the common-case. Convolve the input without any further complications.             output = convolve(inp, kernel)             # Add the biases             if biased:                 biases = self.make_var('biases', [c_o])                 output = tf.nn.bias_add(output, biases)             if relu:                 # ReLU non-linearity                 output = tf.nn.relu(output, name=scope.name)             return output       @layer     def prelu(self, inp, name):         with tf.variable_scope(name):             i = int(inp.get_shape()[-1])             alpha = self.make_var('alpha', shape=(i,))             output = tf.nn.relu(inp) + tf.multiply(alpha, -tf.nn.relu(-inp))         return output       @layer     def max_pool(self, inp, k_h, k_w, s_h, s_w, name, padding='SAME'):         self.validate_padding(padding)         return tf.nn.max_pool(inp,                               ksize=[1, k_h, k_w, 1],                               strides=[1, s_h, s_w, 1],                               padding=padding,                               name=name)       @layer     def fc(self, inp, num_out, name, relu=True):         with tf.variable_scope(name):             input_shape = inp.get_shape()             if input_shape.ndims == 4:                 # The input is spatial. Vectorize it first.                 dim = 1                 for d in input_shape[1:].as_list():                     dim *= int(d)                 feed_in = tf.reshape(inp, [-1, dim])             else:                 feed_in, dim = (inp, input_shape[-1].value)             weights = self.make_var('weights', shape=[dim, num_out])             biases = self.make_var('biases', [num_out])             op = tf.nn.relu_layer if relu else tf.nn.xw_plus_b             fc = op(feed_in, weights, biases, name=name)             return fc         """     Multi dimensional softmax,     refer to https://github.com/tensorflow/tensorflow/issues/210     compute softmax along the dimension of target     the native softmax only supports batch_size x dimension     """     @layer     def softmax(self, target, axis, name=None):         max_axis = tf.reduce_max(target, axis, keepdims=True)         target_exp = tf.exp(target-max_axis)         normalize = tf.reduce_sum(target_exp, axis, keepdims=True)         softmax = tf.div(target_exp, normalize, name)         return softmax       class PNet(Network):     def setup(self):         (self.feed('data') #pylint: disable=no-value-for-parameter, no-member              .conv(3, 3, 10, 1, 1, padding='VALID', relu=False, name='conv1')              .prelu(name='PReLU1')              .max_pool(2, 2, 2, 2, name='pool1')              .conv(3, 3, 16, 1, 1, padding='VALID', relu=False, name='conv2')              .prelu(name='PReLU2')              .conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv3')              .prelu(name='PReLU3')              .conv(1, 1, 2, 1, 1, relu=False, name='conv4-1')              .softmax(3,name='prob1'))           (self.feed('PReLU3') #pylint: disable=no-value-for-parameter              .conv(1, 1, 4, 1, 1, relu=False, name='conv4-2'))           class RNet(Network):     def setup(self):         (self.feed('data') #pylint: disable=no-value-for-parameter, no-member              .conv(3, 3, 28, 1, 1, padding='VALID', relu=False, name='conv1')              .prelu(name='prelu1')              .max_pool(3, 3, 2, 2, name='pool1')              .conv(3, 3, 48, 1, 1, padding='VALID', relu=False, name='conv2')              .prelu(name='prelu2')              .max_pool(3, 3, 2, 2, padding='VALID', name='pool2')              .conv(2, 2, 64, 1, 1, padding='VALID', relu=False, name='conv3')              .prelu(name='prelu3')              .fc(128, relu=False, name='conv4')              .prelu(name='prelu4')              .fc(2, relu=False, name='conv5-1')              .softmax(1,name='prob1'))           (self.feed('prelu4') #pylint: disable=no-value-for-parameter              .fc(4, relu=False, name='conv5-2'))   class ONet(Network):     def setup(self):         (self.feed('data') #pylint: disable=no-value-for-parameter, no-member              .conv(3, 3, 32, 1, 1, padding='VALID', relu=False, name='conv1')              .prelu(name='prelu1')              .max_pool(3, 3, 2, 2, name='pool1')              .conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv2')              .prelu(name='prelu2')              .max_pool(3, 3, 2, 2, padding='VALID', name='pool2')              .conv(3, 3, 64, 1, 1, padding='VALID', relu=False, name='conv3')              .prelu(name='prelu3')              .max_pool(2, 2, 2, 2, name='pool3')              .conv(2, 2, 128, 1, 1, padding='VALID', relu=False, name='conv4')              .prelu(name='prelu4')              .fc(256, relu=False, name='conv5')              .prelu(name='prelu5')              .fc(2, relu=False, name='conv6-1')              .softmax(1, name='prob1'))           (self.feed('prelu5') #pylint: disable=no-value-for-parameter              .fc(4, relu=False, name='conv6-2'))           (self.feed('prelu5') #pylint: disable=no-value-for-parameter              .fc(10, relu=False, name='conv6-3'))   def create_mtcnn(sess, model_path):     if not model_path:         model_path,_ = os.path.split(os.path.realpath(__file__))       with tf.variable_scope('pnet'):         data = tf.placeholder(tf.float32, (None,None,None,3), 'input')         pnet = PNet({'data':data})         pnet.load(os.path.join(model_path, 'det1.npy'), sess)     with tf.variable_scope('rnet'):         data = tf.placeholder(tf.float32, (None,24,24,3), 'input')         rnet = RNet({'data':data})         rnet.load(os.path.join(model_path, 'det2.npy'), sess)     with tf.variable_scope('onet'):         data = tf.placeholder(tf.float32, (None,48,48,3), 'input')         onet = ONet({'data':data})         onet.load(os.path.join(model_path, 'det3.npy'), sess)               pnet_fun = lambda img : sess.run(('pnet/conv4-2/BiasAdd:0', 'pnet/prob1:0'), feed_dict={'pnet/input:0':img})     rnet_fun = lambda img : sess.run(('rnet/conv5-2/conv5-2:0', 'rnet/prob1:0'), feed_dict={'rnet/input:0':img})     onet_fun = lambda img : sess.run(('onet/conv6-2/conv6-2:0', 'onet/conv6-3/conv6-3:0', 'onet/prob1:0'), feed_dict={'onet/input:0':img})     return pnet_fun, rnet_fun, onet_fun   def detect_face(img, minsize, pnet, rnet, onet, threshold, factor):     """Detects faces in an image, and returns bounding boxes and points for them.     img: input image     minsize: minimum faces' size     pnet, rnet, onet: caffemodel     threshold: threshold=[th1, th2, th3], th1-3 are three steps's threshold     factor: the factor used to create a scaling pyramid of face sizes to detect in the image.     """     factor_count=0     total_boxes=np.empty((0,9))     points=np.empty(0)     h=img.shape[0]     w=img.shape[1]     minl=np.amin([h, w])     m=12.0/minsize     minl=minl*m     # create scale pyramid     scales=[]     while minl>=12:         scales += [m*np.power(factor, factor_count)]         minl = minl*factor         factor_count += 1       # first stage     for scale in scales:         hs=int(np.ceil(h*scale))         ws=int(np.ceil(w*scale))         im_data = imresample(img, (hs, ws))         im_data = (im_data-127.5)*0.0078125         img_x = np.expand_dims(im_data, 0)         img_y = np.transpose(img_x, (0,2,1,3))         out = pnet(img_y)         out0 = np.transpose(out[0], (0,2,1,3))         out1 = np.transpose(out[1], (0,2,1,3))                   boxes, _ = generateBoundingBox(out1[0,:,:,1].copy(), out0[0,:,:,:].copy(), scale, threshold[0])                   # inter-scale nms         pick = nms(boxes.copy(), 0.5, 'Union')         if boxes.size>0 and pick.size>0:             boxes = boxes[pick,:]             total_boxes = np.append(total_boxes, boxes, axis=0)       numbox = total_boxes.shape[0]     if numbox>0:         pick = nms(total_boxes.copy(), 0.7, 'Union')         total_boxes = total_boxes[pick,:]         regw = total_boxes[:,2]-total_boxes[:,0]         regh = total_boxes[:,3]-total_boxes[:,1]         qq1 = total_boxes[:,0]+total_boxes[:,5]*regw         qq2 = total_boxes[:,1]+total_boxes[:,6]*regh         qq3 = total_boxes[:,2]+total_boxes[:,7]*regw         qq4 = total_boxes[:,3]+total_boxes[:,8]*regh         total_boxes = np.transpose(np.vstack([qq1, qq2, qq3, qq4, total_boxes[:,4]]))         total_boxes = rerec(total_boxes.copy())         total_boxes[:,0:4] = np.fix(total_boxes[:,0:4]).astype(np.int32)         dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)       numbox = total_boxes.shape[0]     if numbox>0:         # second stage         tempimg = np.zeros((24,24,3,numbox))         for k in range(0,numbox):             tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))             tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]             if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:                 tempimg[:,:,:,k] = imresample(tmp, (24, 24))             else:                 return np.empty()         tempimg = (tempimg-127.5)*0.0078125         tempimg1 = np.transpose(tempimg, (3,1,0,2))         out = rnet(tempimg1)         out0 = np.transpose(out[0])         out1 = np.transpose(out[1])         score = out1[1,:]         ipass = np.where(score>threshold[1])         total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])         mv = out0[:,ipass[0]]         if total_boxes.shape[0]>0:             pick = nms(total_boxes, 0.7, 'Union')             total_boxes = total_boxes[pick,:]             total_boxes = bbreg(total_boxes.copy(), np.transpose(mv[:,pick]))             total_boxes = rerec(total_boxes.copy())       numbox = total_boxes.shape[0]     if numbox>0:         # third stage         total_boxes = np.fix(total_boxes).astype(np.int32)         dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(total_boxes.copy(), w, h)         tempimg = np.zeros((48,48,3,numbox))         for k in range(0,numbox):             tmp = np.zeros((int(tmph[k]),int(tmpw[k]),3))             tmp[dy[k]-1:edy[k],dx[k]-1:edx[k],:] = img[y[k]-1:ey[k],x[k]-1:ex[k],:]             if tmp.shape[0]>0 and tmp.shape[1]>0 or tmp.shape[0]==0 and tmp.shape[1]==0:                 tempimg[:,:,:,k] = imresample(tmp, (48, 48))             else:                 return np.empty()         tempimg = (tempimg-127.5)*0.0078125         tempimg1 = np.transpose(tempimg, (3,1,0,2))         out = onet(tempimg1)         out0 = np.transpose(out[0])         out1 = np.transpose(out[1])         out2 = np.transpose(out[2])         score = out2[1,:]         points = out1         ipass = np.where(score>threshold[2])         points = points[:,ipass[0]]         total_boxes = np.hstack([total_boxes[ipass[0],0:4].copy(), np.expand_dims(score[ipass].copy(),1)])         mv = out0[:,ipass[0]]           w = total_boxes[:,2]-total_boxes[:,0]+1         h = total_boxes[:,3]-total_boxes[:,1]+1         points[0:5,:] = np.tile(w,(5, 1))*points[0:5,:] + np.tile(total_boxes[:,0],(5, 1))-1         points[5:10,:] = np.tile(h,(5, 1))*points[5:10,:] + np.tile(total_boxes[:,1],(5, 1))-1         if total_boxes.shape[0]>0:             total_boxes = bbreg(total_boxes.copy(), np.transpose(mv))             pick = nms(total_boxes.copy(), 0.7, 'Min')             total_boxes = total_boxes[pick,:]             points = points[:,pick]                       return total_boxes, points     def bulk_detect_face(images, detection_window_size_ratio, pnet, rnet, onet, threshold, factor):     """Detects faces in a list of images     images: list containing input images     detection_window_size_ratio: ratio of minimum face size to smallest image dimension     pnet, rnet, onet: caffemodel     threshold: threshold=[th1 th2 th3], th1-3 are three steps's threshold [0-1]     factor: the factor used to create a scaling pyramid of face sizes to detect in the image.     """     all_scales = [None] * len(images)     images_with_boxes = [None] * len(images)       for i in range(len(images)):         images_with_boxes[i] = {'total_boxes': np.empty((0, 9))}       # create scale pyramid     for index, img in enumerate(images):         all_scales[index] = []         h = img.shape[0]         w = img.shape[1]         minsize = int(detection_window_size_ratio * np.minimum(w, h))         factor_count = 0         minl = np.amin([h, w])         if minsize <= 12:             minsize = 12           m = 12.0 / minsize         minl = minl * m         while minl >= 12:             all_scales[index].append(m * np.power(factor, factor_count))             minl = minl * factor             factor_count += 1       # # # # # # # # # # # # #     # first stage - fast proposal network (pnet) to obtain face candidates     # # # # # # # # # # # # #       images_obj_per_resolution = {}       # TODO: use some type of rounding to number module 8 to increase probability that pyramid images will have the same resolution across input images       for index, scales in enumerate(all_scales):         h = images[index].shape[0]         w = images[index].shape[1]           for scale in scales:             hs = int(np.ceil(h * scale))             ws = int(np.ceil(w * scale))               if (ws, hs) not in images_obj_per_resolution:                 images_obj_per_resolution[(ws, hs)] = []               im_data = imresample(images[index], (hs, ws))             im_data = (im_data - 127.5) * 0.0078125             img_y = np.transpose(im_data, (1, 0, 2))  # caffe uses different dimensions ordering             images_obj_per_resolution[(ws, hs)].append({'scale': scale, 'image': img_y, 'index': index})       for resolution in images_obj_per_resolution:         images_per_resolution = [i['image'] for i in images_obj_per_resolution[resolution]]         outs = pnet(images_per_resolution)           for index in range(len(outs[0])):             scale = images_obj_per_resolution[resolution][index]['scale']             image_index = images_obj_per_resolution[resolution][index]['index']             out0 = np.transpose(outs[0][index], (1, 0, 2))             out1 = np.transpose(outs[1][index], (1, 0, 2))               boxes, _ = generateBoundingBox(out1[:, :, 1].copy(), out0[:, :, :].copy(), scale, threshold[0])               # inter-scale nms             pick = nms(boxes.copy(), 0.5, 'Union')             if boxes.size > 0 and pick.size > 0:                 boxes = boxes[pick, :]                 images_with_boxes[image_index]['total_boxes'] = np.append(images_with_boxes[image_index]['total_boxes'],                                                                           boxes,                                                                           axis=0)       for index, image_obj in enumerate(images_with_boxes):         numbox = image_obj['total_boxes'].shape[0]         if numbox > 0:             h = images[index].shape[0]             w = images[index].shape[1]             pick = nms(image_obj['total_boxes'].copy(), 0.7, 'Union')             image_obj['total_boxes'] = image_obj['total_boxes'][pick, :]             regw = image_obj['total_boxes'][:, 2] - image_obj['total_boxes'][:, 0]             regh = image_obj['total_boxes'][:, 3] - image_obj['total_boxes'][:, 1]             qq1 = image_obj['total_boxes'][:, 0] + image_obj['total_boxes'][:, 5] * regw             qq2 = image_obj['total_boxes'][:, 1] + image_obj['total_boxes'][:, 6] * regh             qq3 = image_obj['total_boxes'][:, 2] + image_obj['total_boxes'][:, 7] * regw             qq4 = image_obj['total_boxes'][:, 3] + image_obj['total_boxes'][:, 8] * regh             image_obj['total_boxes'] = np.transpose(np.vstack([qq1, qq2, qq3, qq4, image_obj['total_boxes'][:, 4]]))             image_obj['total_boxes'] = rerec(image_obj['total_boxes'].copy())             image_obj['total_boxes'][:, 0:4] = np.fix(image_obj['total_boxes'][:, 0:4]).astype(np.int32)             dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(image_obj['total_boxes'].copy(), w, h)               numbox = image_obj['total_boxes'].shape[0]             tempimg = np.zeros((24, 24, 3, numbox))               if numbox > 0:                 for k in range(0, numbox):                     tmp = np.zeros((int(tmph[k]), int(tmpw[k]), 3))                     tmp[dy[k] - 1:edy[k], dx[k] - 1:edx[k], :] = images[index][y[k] - 1:ey[k], x[k] - 1:ex[k], :]                     if tmp.shape[0] > 0 and tmp.shape[1] > 0 or tmp.shape[0] == 0 and tmp.shape[1] == 0:                         tempimg[:, :, :, k] = imresample(tmp, (24, 24))                     else:                         return np.empty()                   tempimg = (tempimg - 127.5) * 0.0078125                 image_obj['rnet_input'] = np.transpose(tempimg, (3, 1, 0, 2))       # # # # # # # # # # # # #     # second stage - refinement of face candidates with rnet     # # # # # # # # # # # # #       bulk_rnet_input = np.empty((0, 24, 24, 3))     for index, image_obj in enumerate(images_with_boxes):         if 'rnet_input' in image_obj:             bulk_rnet_input = np.append(bulk_rnet_input, image_obj['rnet_input'], axis=0)       out = rnet(bulk_rnet_input)     out0 = np.transpose(out[0])     out1 = np.transpose(out[1])     score = out1[1, :]       i = 0     for index, image_obj in enumerate(images_with_boxes):         if 'rnet_input' not in image_obj:             continue           rnet_input_count = image_obj['rnet_input'].shape[0]         score_per_image = score[i:i + rnet_input_count]         out0_per_image = out0[:, i:i + rnet_input_count]           ipass = np.where(score_per_image > threshold[1])         image_obj['total_boxes'] = np.hstack([image_obj['total_boxes'][ipass[0], 0:4].copy(),                                               np.expand_dims(score_per_image[ipass].copy(), 1)])           mv = out0_per_image[:, ipass[0]]           if image_obj['total_boxes'].shape[0] > 0:             h = images[index].shape[0]             w = images[index].shape[1]             pick = nms(image_obj['total_boxes'], 0.7, 'Union')             image_obj['total_boxes'] = image_obj['total_boxes'][pick, :]             image_obj['total_boxes'] = bbreg(image_obj['total_boxes'].copy(), np.transpose(mv[:, pick]))             image_obj['total_boxes'] = rerec(image_obj['total_boxes'].copy())               numbox = image_obj['total_boxes'].shape[0]               if numbox > 0:                 tempimg = np.zeros((48, 48, 3, numbox))                 image_obj['total_boxes'] = np.fix(image_obj['total_boxes']).astype(np.int32)                 dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph = pad(image_obj['total_boxes'].copy(), w, h)                   for k in range(0, numbox):                     tmp = np.zeros((int(tmph[k]), int(tmpw[k]), 3))                     tmp[dy[k] - 1:edy[k], dx[k] - 1:edx[k], :] = images[index][y[k] - 1:ey[k], x[k] - 1:ex[k], :]                     if tmp.shape[0] > 0 and tmp.shape[1] > 0 or tmp.shape[0] == 0 and tmp.shape[1] == 0:                         tempimg[:, :, :, k] = imresample(tmp, (48, 48))                     else:                         return np.empty()                 tempimg = (tempimg - 127.5) * 0.0078125                 image_obj['onet_input'] = np.transpose(tempimg, (3, 1, 0, 2))           i += rnet_input_count       # # # # # # # # # # # # #     # third stage - further refinement and facial landmarks positions with onet     # # # # # # # # # # # # #       bulk_onet_input = np.empty((0, 48, 48, 3))     for index, image_obj in enumerate(images_with_boxes):         if 'onet_input' in image_obj:             bulk_onet_input = np.append(bulk_onet_input, image_obj['onet_input'], axis=0)       out = onet(bulk_onet_input)       out0 = np.transpose(out[0])     out1 = np.transpose(out[1])     out2 = np.transpose(out[2])     score = out2[1, :]     points = out1       i = 0     ret = []     for index, image_obj in enumerate(images_with_boxes):         if 'onet_input' not in image_obj:             ret.append(None)             continue           onet_input_count = image_obj['onet_input'].shape[0]           out0_per_image = out0[:, i:i + onet_input_count]         score_per_image = score[i:i + onet_input_count]         points_per_image = points[:, i:i + onet_input_count]           ipass = np.where(score_per_image > threshold[2])         points_per_image = points_per_image[:, ipass[0]]           image_obj['total_boxes'] = np.hstack([image_obj['total_boxes'][ipass[0], 0:4].copy(),                                               np.expand_dims(score_per_image[ipass].copy(), 1)])         mv = out0_per_image[:, ipass[0]]           w = image_obj['total_boxes'][:, 2] - image_obj['total_boxes'][:, 0] + 1         h = image_obj['total_boxes'][:, 3] - image_obj['total_boxes'][:, 1] + 1         points_per_image[0:5, :] = np.tile(w, (5, 1)) * points_per_image[0:5, :] + np.tile(             image_obj['total_boxes'][:, 0], (5, 1)) - 1         points_per_image[5:10, :] = np.tile(h, (5, 1)) * points_per_image[5:10, :] + np.tile(             image_obj['total_boxes'][:, 1], (5, 1)) - 1           if image_obj['total_boxes'].shape[0] > 0:             image_obj['total_boxes'] = bbreg(image_obj['total_boxes'].copy(), np.transpose(mv))             pick = nms(image_obj['total_boxes'].copy(), 0.7, 'Min')             image_obj['total_boxes'] = image_obj['total_boxes'][pick, :]             points_per_image = points_per_image[:, pick]               ret.append((image_obj['total_boxes'], points_per_image))         else:             ret.append(None)           i += onet_input_count       return ret     # function [boundingbox] = bbreg(boundingbox,reg) def bbreg(boundingbox,reg):     """Calibrate bounding boxes"""     if reg.shape[1]==1:         reg = np.reshape(reg, (reg.shape[2], reg.shape[3]))       w = boundingbox[:,2]-boundingbox[:,0]+1     h = boundingbox[:,3]-boundingbox[:,1]+1     b1 = boundingbox[:,0]+reg[:,0]*w     b2 = boundingbox[:,1]+reg[:,1]*h     b3 = boundingbox[:,2]+reg[:,2]*w     b4 = boundingbox[:,3]+reg[:,3]*h     boundingbox[:,0:4] = np.transpose(np.vstack([b1, b2, b3, b4 ]))     return boundingbox    def generateBoundingBox(imap, reg, scale, t):     """Use heatmap to generate bounding boxes"""     stride=2     cellsize=12       imap = np.transpose(imap)     dx1 = np.transpose(reg[:,:,0])     dy1 = np.transpose(reg[:,:,1])     dx2 = np.transpose(reg[:,:,2])     dy2 = np.transpose(reg[:,:,3])     y, x = np.where(imap >= t)     if y.shape[0]==1:         dx1 = np.flipud(dx1)         dy1 = np.flipud(dy1)         dx2 = np.flipud(dx2)         dy2 = np.flipud(dy2)     score = imap[(y,x)]     reg = np.transpose(np.vstack([ dx1[(y,x)], dy1[(y,x)], dx2[(y,x)], dy2[(y,x)] ]))     if reg.size==0:         reg = np.empty((0,3))     bb = np.transpose(np.vstack([y,x]))     q1 = np.fix((stride*bb+1)/scale)     q2 = np.fix((stride*bb+cellsize-1+1)/scale)     boundingbox = np.hstack([q1, q2, np.expand_dims(score,1), reg])     return boundingbox, reg    # function pick = nms(boxes,threshold,type) def nms(boxes, threshold, method):     if boxes.size==0:         return np.empty((0,3))     x1 = boxes[:,0]     y1 = boxes[:,1]     x2 = boxes[:,2]     y2 = boxes[:,3]     s = boxes[:,4]     area = (x2-x1+1) * (y2-y1+1)     I = np.argsort(s)     pick = np.zeros_like(s, dtype=np.int16)     counter = 0     while I.size>0:         i = I[-1]         pick[counter] = i         counter += 1         idx = I[0:-1]         xx1 = np.maximum(x1[i], x1[idx])         yy1 = np.maximum(y1[i], y1[idx])         xx2 = np.minimum(x2[i], x2[idx])         yy2 = np.minimum(y2[i], y2[idx])         w = np.maximum(0.0, xx2-xx1+1)         h = np.maximum(0.0, yy2-yy1+1)         inter = w * h         if method is 'Min':             o = inter / np.minimum(area[i], area[idx])         else:             o = inter / (area[i] + area[idx] - inter)         I = I[np.where(o<=threshold)]     pick = pick[0:counter]     return pick   # function [dy edy dx edx y ey x ex tmpw tmph] = pad(total_boxes,w,h) def pad(total_boxes, w, h):     """Compute the padding coordinates (pad the bounding boxes to square)"""     tmpw = (total_boxes[:,2]-total_boxes[:,0]+1).astype(np.int32)     tmph = (total_boxes[:,3]-total_boxes[:,1]+1).astype(np.int32)     numbox = total_boxes.shape[0]       dx = np.ones((numbox), dtype=np.int32)     dy = np.ones((numbox), dtype=np.int32)     edx = tmpw.copy().astype(np.int32)     edy = tmph.copy().astype(np.int32)       x = total_boxes[:,0].copy().astype(np.int32)     y = total_boxes[:,1].copy().astype(np.int32)     ex = total_boxes[:,2].copy().astype(np.int32)     ey = total_boxes[:,3].copy().astype(np.int32)       tmp = np.where(ex>w)     edx.flat[tmp] = np.expand_dims(-ex[tmp]+w+tmpw[tmp],1)     ex[tmp] = w           tmp = np.where(ey>h)     edy.flat[tmp] = np.expand_dims(-ey[tmp]+h+tmph[tmp],1)     ey[tmp] = h       tmp = np.where(x<1)     dx.flat[tmp] = np.expand_dims(2-x[tmp],1)     x[tmp] = 1       tmp = np.where(y<1)     dy.flat[tmp] = np.expand_dims(2-y[tmp],1)     y[tmp] = 1           return dy, edy, dx, edx, y, ey, x, ex, tmpw, tmph   # function [bboxA] = rerec(bboxA) def rerec(bboxA):     """Convert bboxA to square."""     h = bboxA[:,3]-bboxA[:,1]     w = bboxA[:,2]-bboxA[:,0]     l = np.maximum(w, h)     bboxA[:,0] = bboxA[:,0]+w*0.5-l*0.5     bboxA[:,1] = bboxA[:,1]+h*0.5-l*0.5     bboxA[:,2:4] = bboxA[:,0:2] + np.transpose(np.tile(l,(2,1)))     return bboxA   def imresample(img, sz):     im_data = cv2.resize(img, (sz[1], sz[0]), interpolation=cv2.INTER_AREA) #@UndefinedVariable     return im_data       # This method is kept for debugging purpose #     h=img.shape[0] #     w=img.shape[1] #     hs, ws = sz #     dx = float(w) / ws #     dy = float(h) / hs #     im_data = np.zeros((hs,ws,3)) #     for a1 in range(0,hs): #         for a2 in range(0,ws): #             for a3 in range(0,3): #                 im_data[a1,a2,a3] = img[int(floor(a1*dy)),int(floor(a2*dx)),a3] #     return im_data

　　

 

 

参考：

1 https://zhuanlan.zhihu.com/p/25025596

2 https://link.zhihu.com/?target=https%3A//github.com/shanren7/real_time_face_recognition

3 https://blog.csdn.net/mr_evanchen/article/details/77650883

4 https://github.com/ShyBigBoy/face-detection-mtcnn