Research On Shoe Pattern Recognition Method Based On Convolutional Neural Network

The method,so-called "monitoring + shoeprint",plays a pivotal role in criminal investigation.The basic principle is to infer the shoes worn by the suspect according to the shoeprints at the crime scene,and then retrieve the shoeprints of the suspect from the surrounding monitoring video to obtain the image information of the criminal suspect,so as to provide effective clues for the investigation of the case.However,"monitoring + shoeprint" method remains to rely on manual work with relatively low efficiency.Therefore,we desperately need an automated identifying method of shoe pattern which is based on the image of shoes obtained through analyzing shoeprints in crime scene and could automatically match with the individual in monitoring video.Hence,it provided a quick and effective method for tracking the suspectAiming at the problem of the low degree of automation of the "monitoring + shoeprint" technique,this paper concentrates on the shoe's pattern recognition method based on convolutional neural network to realize the automatic recognition method of suspicious shoe patterns.First of all,a dataset about shoe's pattern did not exist.So,we collected videos by simulating the surveillance video around the crime scene,which were processed frame by frame.And then we designed cutting codes to simulate the automatic retrieval and cutting process of shoe samples.The video frames were manually cut and bilinear interpolation was used to normalize them.We established a shoe sample database containing 50 different types of shoes and 160 231 pictures.Secondly,this paper used the Caffe framework.Based on two classic neural network structures named LeNet and DeepID,this paper established a series of novel networks by deepening the layers of the two classic networks,and replacing the Relu activation function with the PRelu activation function.There were five types of network structures for training and testing on shoe's pattern data.The results show that the DeepIDadd + PRelu network achieved the best accuracy of 97.73%,which is 3.37% higher than the accuracy of the DeepID network.In addition,a robustness test was performed on the optimal network while similarity measures between different shoe samples were estimated for shoe shape verification.At the same time,analysis of misidentified pictures was performed to facilitate further research work.The results show that the convolutional neural network has good recognition effect and strong robustness in recognition of shoe pattern which provides a new way for shoe pattern recognition.Then,because the model is large and difficult to practically apply,this paper proposes a lightweight convolutional neural network structure based on DeepIDadd + PRelu named SPRnet.We introduced depthwise separable convolution to replace original convolution,and introduced global average pooling to replace the first full connection layer.As consequence,thesize reduced to 0.5M,and the single-sheet recognition only spent 60 ms.The results show that the SPR-net greatly reduced the model size and improved the speed.Finally,we can rank the similarities by calculating the cosine similarity of extracted features.This paper designed the SPR shoe pattern recognition system to display the top six shoes with similarity.