开源KerasFasterRCNN模型介绍及代码结构解析

开源Keras Faster RCNN 模型地址&＃xff1a;https://github.com/jinfagang/keras_frcnn

• 1、由于该源代码由keras单独库编写所以需要下载keras,必须是2.0.3版本

pip install keras&＃61;&＃61;2.0.3


• 2、该源码读取图片以及处理图片标记图片工具使用opecv需要安装

pip install opencv-python

2. keras Faster RCNN代码结构

源码组成结构&＃xff1a;

• detector&＃xff1a;FasterRCNNDetector目标检测器代码

img_input &＃61; Input(shape&＃61;input_shape_img)roi_input &＃61; Input(shape&＃61;(None, 4))# define the base network (resnet here, can be VGG, Inception, etc)shared_layers &＃61; nn.nn_base(img_input, trainable&＃61;True)# define the RPN, built on the base layersnum_anchors &＃61; len(cfg.anchor_box_scales) * len(cfg.anchor_box_ratios)rpn &＃61; nn.rpn(shared_layers, num_anchors)classifier &＃61; nn.classifier(shared_layers, roi_input, cfg.num_rois, nb_classes&＃61;len(classes_count), trainable&＃61;True)model_rpn &＃61; Model(img_input, rpn[:2])model_classifier &＃61; Model([img_input, roi_input], classifier)# this is a model that holds both the RPN and the classifier, used to load/save weights for the modelsmodel_all &＃61; Model([img_input, roi_input], rpn[:2] &＃43; classifier)try:print(&＃39;loading weights from {}&＃39;.format(cfg.base_net_weights))model_rpn.load_weights(cfg.model_path, by_name&＃61;True)model_classifier.load_weights(cfg.model_path, by_name&＃61;True)except Exception as e:print(e)print(&＃39;Could not load pretrained model weights. Weights can be found in the keras application folder &＃39;&＃39;https://github.com/fchollet/keras/tree/master/keras/applications&＃39;)optimizer &＃61; Adam(lr&＃61;1e-5)optimizer_classifier &＃61; Adam(lr&＃61;1e-5)model_rpn.compile(optimizer&＃61;optimizer,loss&＃61;[losses_fn.rpn_loss_cls(num_anchors), losses_fn.rpn_loss_regr(num_anchors)])model_classifier.compile(optimizer&＃61;optimizer_classifier,loss&＃61;[losses_fn.class_loss_cls, losses_fn.class_loss_regr(len(classes_count) - 1)],metrics&＃61;{&＃39;dense_class_{}&＃39;.format(len(classes_count)): &＃39;accuracy&＃39;})model_all.compile(optimizer&＃61;&＃39;sgd&＃39;, loss&＃61;&＃39;mae&＃39;)

4. RPN 与 classifier定义

• RPN结构

def rpn(base_layers, num_anchors):x &＃61; Convolution2D(512, (3, 3), padding&＃61;&＃39;same&＃39;, activation&＃61;&＃39;relu&＃39;, kernel_initializer&＃61;&＃39;normal&＃39;, name&＃61;&＃39;rpn_conv1&＃39;)(base_layers)x_class &＃61; Convolution2D(num_anchors, (1, 1), activation&＃61;&＃39;sigmoid&＃39;, kernel_initializer&＃61;&＃39;uniform&＃39;,name&＃61;&＃39;rpn_out_class&＃39;)(x)x_regr &＃61; Convolution2D(num_anchors * 4, (1, 1), activation&＃61;&＃39;linear&＃39;, kernel_initializer&＃61;&＃39;zero&＃39;,name&＃61;&＃39;rpn_out_regress&＃39;)(x)return [x_class, x_regr, base_layers]


• classifier结构

def classifier(base_layers, input_rois, num_rois, nb_classes&＃61;21, trainable&＃61;False):# compile times on theano tend to be very high, so we use smaller ROI pooling regions to workaroundif K.backend() &＃61;&＃61; &＃39;tensorflow&＃39;:pooling_regions &＃61; 14input_shape &＃61; (num_rois, 14, 14, 1024)elif K.backend() &＃61;&＃61; &＃39;theano&＃39;:pooling_regions &＃61; 7input_shape &＃61; (num_rois, 1024, 7, 7)# ROI pooling计算定义out_roi_pool &＃61; RoiPoolingConv(pooling_regions, num_rois)([base_layers, input_rois])out &＃61; classifier_layers(out_roi_pool, input_shape&＃61;input_shape, trainable&＃61;True)out &＃61; TimeDistributed(Flatten())(out)# 分类out_class &＃61; TimeDistributed(Dense(nb_classes, activation&＃61;&＃39;softmax&＃39;, kernel_initializer&＃61;&＃39;zero&＃39;),name&＃61;&＃39;dense_class_{}&＃39;.format(nb_classes))(out)# note: no regression target for bg class# 回归out_regr &＃61; TimeDistributed(Dense(4 * (nb_classes - 1), activation&＃61;&＃39;linear&＃39;, kernel_initializer&＃61;&＃39;zero&＃39;),name&＃61;&＃39;dense_regress_{}&＃39;.format(nb_classes))(out)return [out_class, out_regr]

5. data_generators.py:传递图像参数&＃xff0c;增广配置参数&＃xff0c;是否进行图像增广

• IoU计算&＃xff1a;

from __future__ import absolute_import
import numpy as np
import cv2
import random
import copy
from . import data_augment
import threading
import itertools#并集
def union(au, bu, area_intersection):area_a &＃61; (au[2] - au[0]) * (au[3] - au[1])area_b &＃61; (bu[2] - bu[0]) * (bu[3] - bu[1])area_union &＃61; area_a &＃43; area_b - area_intersectionreturn area_union#交集
def intersection(ai, bi):x &＃61; max(ai[0], bi[0])y &＃61; max(ai[1], bi[1])w &＃61; min(ai[2], bi[2]) - xh &＃61; min(ai[3], bi[3]) - yif w < 0 or h < 0:return 0return w*h#交并比
def iou(a, b):# a and b should be (x1,y1,x2,y2)if a[0] >&＃61; a[2] or a[1] >&＃61; a[3] or b[0] >&＃61; b[2] or b[1] >&＃61; b[3]:return 0.0area_i &＃61; intersection(a, b)area_u &＃61; union(a, b, area_i)return float(area_i) / float(area_u &＃43; 1e-6)

6. 损失计算 losses.py

• rpn的损失回归和分类

def rpn_loss_regr(num_anchors):def rpn_loss_regr_fixed_num(y_true, y_pred):if K.image_dim_ordering() &＃61;&＃61; &＃39;th&＃39;:x &＃61; y_true[:, 4 * num_anchors:, :, :] - y_predx_abs &＃61; K.abs(x)x_bool &＃61; K.less_equal(x_abs, 1.0)return lambda_rpn_regr * K.sum(y_true[:, :4 * num_anchors, :, :] * (x_bool * (0.5 * x * x) &＃43; (1 - x_bool) * (x_abs - 0.5))) / K.sum(epsilon &＃43; y_true[:, :4 * num_anchors, :, :])else:x &＃61; y_true[:, :, :, 4 * num_anchors:] - y_predx_abs &＃61; K.abs(x)x_bool &＃61; K.cast(K.less_equal(x_abs, 1.0), tf.float32)return lambda_rpn_regr * K.sum(y_true[:, :, :, :4 * num_anchors] * (x_bool * (0.5 * x * x) &＃43; (1 - x_bool) * (x_abs - 0.5))) / K.sum(epsilon &＃43; y_true[:, :, :, :4 * num_anchors])return rpn_loss_regr_fixed_numdef rpn_loss_cls(num_anchors):def rpn_loss_cls_fixed_num(y_true, y_pred):if K.image_dim_ordering() &＃61;&＃61; &＃39;tf&＃39;:return lambda_rpn_class * K.sum(y_true[:, :, :, :num_anchors] * K.binary_crossentropy(y_pred[:, :, :, :], y_true[:, :, :, num_anchors:])) / K.sum(epsilon &＃43; y_true[:, :, :, :num_anchors])else:return lambda_rpn_class * K.sum(y_true[:, :num_anchors, :, :] * K.binary_crossentropy(y_pred[:, :, :, :], y_true[:, num_anchors:, :, :])) / K.sum(epsilon &＃43; y_true[:, :num_anchors, :, :])return rpn_loss_cls_fixed_num


• fastrcnn的分类和回归

def class_loss_regr(num_classes):def class_loss_regr_fixed_num(y_true, y_pred):x &＃61; y_true[:, :, 4*num_classes:] - y_predx_abs &＃61; K.abs(x)x_bool &＃61; K.cast(K.less_equal(x_abs, 1.0), &＃39;float32&＃39;)return lambda_cls_regr * K.sum(y_true[:, :, :4*num_classes] * (x_bool * (0.5 * x * x) &＃43; (1 - x_bool) * (x_abs - 0.5))) / K.sum(epsilon &＃43; y_true[:, :, :4*num_classes])return class_loss_regr_fixed_numdef class_loss_cls(y_true, y_pred):return lambda_cls_class * K.mean(categorical_crossentropy(y_true[0, :, :], y_pred[0, :, :]))


加油!

感谢!

努力!