基于Faster R-CNN的服装目标检测改进方法  

Improved Method for Garment Target Detection Based on Faster R-CNN

作　　者：陈双 何利力 CHEN Shuang;HE Li-li(School of Information,Zhejiang Sci-tec University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学信息学院,浙江杭州310018

出　　处：《软件导刊》2020年第4期42-45,共4页Software Guide

基　　金：浙江省科技厅(重大)项目(2015C03001)。

摘　　要：为了满足近年来在服装电子商务高速发展背景下急剧增长的服装图像分类与定位需求,实现对服装图像的目标检测,提出基于Faster R-CNN的服装目标检测改进方法。借助残差神经网络的特征提取能力进行服装图像特征提取,采用RPN网络进行服装候选区域生成,经过RoI兴趣区域池化后接入分类层与回归层,调整网络结构,融合服装图像的HOG底层特征,从而针对性地实现对服装图像的目标检测。实验结果表明,该方法构建模型的平均准确率为0.902,运行速度为8.9帧/秒,具有一定的实用价值。In order to meet the demand for the classification and positioning of clothing images under the rapid development of clothing e-commerce in recent years,the target detection of clothing images is realized.An improved method of garment target detection based on Faster R-CNN is proposed.The feature extraction ability of residual neural network is used to extract the features of clothing images.The RPN network is used to generate clothing candidate regions.After the pool of RoI interest regions,the classification layer is accessed.With the regression layer,the network structure is adjusted,and the underlying features of the HOG of the clothing image are integrated to achieve the pertinence of the clothing,and the target detection of the clothing image is realized.The experimental results show that the model constructed by this method has an average accuracy of 0.902 and an operating speed of 8.9 frames per second,which proves the method has social practical value and academic research significance.

关 键 词：服装图像 深度学习 目标检测 残差神经网络 边缘梯度直方图特征 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]