“监控+鞋印”是目前公安机关刑事侦查的重要技战法,其基本原理是依据犯罪现场鞋印推断嫌疑人所穿鞋型,然后到周边监控视频中检索嫌疑鞋型。针对“监控+鞋印”技战法自动化程度低下的问题,提出一种基于卷积神经网络的鞋型识别方法,实现对嫌疑鞋型的自动识别。根据鞋型识别独有特点,在DeepID的基础上设计卷积神经网络框架,并构建鞋型样本数据库(50双鞋型样本,共计160231幅图像)。运用Caffe框架结合不同网络模型对鞋型图像数据进行训练和测试,实验设计的初始网络结构由两层卷积、两层池化、两层全连接组成。实验比对了不同的第一层全连接层输出元素数目对网络性能与训练效率的影响,又在不改变输出特征图大小的情况下比对了不同网络深度的实验结果,在优化模型的基础上引用重叠池化得到实验最优网络模型。实验结果表明,卷积神经网络对于鞋型有很好的识别效果,识别精度值最高达96.06%,为鞋型识别提供了一种新的途径。

Criminal investigation is often conducted based on the surveillance video footage and crime-scene shoeprint identification. The basic principle of this method is to infer the type of shoe worn by the suspect based on the shoeprints identified at the crime scene and to subsequently search for the shoe in the surveillance video footage. To solve the problem of low automation associated with this criminal investigation method, a new shoe type recognition method using a convolutional neural network has been proposed in this study. According to the unique characteristics of shoe type recognition, a framework of convolutional neural network is designed on the basis of DeepID, and a shoe database containing 50 pairs of shoes and 160231 images is constructed. The experiments are conducted based on the Caffe framework using different network models. Initially, the network structure comprises two convolution layers, two pooling layers, and two full connection layers. Further, experiments are conducted to compare the effects of the number of output elements in the first layer of two full connection layers on the network performance and training efficiency, and the experimental results of different network depths are compared without changing the size of the output feature graph as well. Based on the optimization model, the optimal network model is obtained by using overlapping pooling. The experimental results denote that the proposed method achieves an excellent performance, with an accuracy of 96.06%. Therefore, the proposed method can be considered to be a promising new method for shoe type recognition.