基于改进残差网络和数据增强的鞋型识别算法

Shoe type recognition algorithm based on improved residual network and data augmentation

在监控视频中搜索犯罪现场出现的嫌疑鞋型是目前侦破案件的重要手段,针对该手段自动化程度低、人工筛选易漏查等问题,提出一种基于改进深度残差网络和数据增强的鞋型识别算法。为增强网络特征提取能力,对深度残差网络进行了研究,在不增加任何参数量的前提下改进瓶颈结构,提升算法精度;针对瓶颈结构中下采样操作存在的问题,改进下采样模块,缓解网络下采样时信息丢失问题;引入Mixup和光学变换数据增强算法,建立数据之间的线性关系,丰富数据的多样性,进而增强网络模型的鲁棒性;最后,采取中心损失函数和Softmax损失函数联合训练的方法,使训练数据达到更好的聚类效果。为验证所提算法的有效性和实用性,在多背景鞋型数据集上对所提算法进行测试,测试结果表明,所提算法mAP、Rank-1精度分别达到66.83%、86.77%,可以有效提高网络识别精度。

It is an important means to search the suspected shoes in the surveillance video at the scene of crime.Aiming at the problems of low automation and manual screening,this paper proposes a shoe recognition algorithm based on improved deep residual network and data augmentation.In order to enhance the ability of network feature extraction,the deep residual network is studied.The bottleneck structure is improved without adding any parameters to improve the accuracy of the algorithm.Aiming at the problem of down sampling operation in bottleneck structure,the down sampling module is improved to alleviate the problem of information loss in network down sampling.Mixup and optical transform data augmentation algorithm are introduced to establish the linear relationship between data,enrich the diversity of data,and enhance the robustness of network model.Finally,the combined training method of center loss function and Softmax loss function is adopted to make the training data achieve better clustering effect.In order to verify the effectiveness and practicability of the proposed algorithm,the proposed algorithm is tested on multi background shoe data sets.The test results show that the accuracy of mAP and Rank-1 of the proposed algorithm is 66.83%and 86.77%respectively,which can effectively improve the accuracy of network recognition.