Research On High Precision And Fast Commodity Detection Algorithm Based On Deep Neural Network

In recent years,with the rapid development of big data,deep network architecture and GPU,many deep learning technologies have been applied to the field of unmanned retail.Therefore,it is of high research and application value to use deep learning technology to realize rapid detection and accurate identification of unmanned supermarket products.Based on this research,this paper proposes a high precision and fast commodity detection algorithm based on deep neural network,in order to reduce operating costs and improve operating efficiency.In view of the large variety of commodities and the presence of multi-scale commodities,a commodity detection algorithm based on multi-scale feature fusion was proposed based on the two-stage detection model Faster-RCNN,which was used to improve the feature extraction ability of small targets and improve the overall detection accuracy of commodities.Moreover,on the basis of multi-scale feature fusion,a commodity detection algorithm based on depth separable convolution is proposed to accelerate the detection speed.These two methods are trained and tested on open RPC data sets.Specific research work is as follows:In order to solve the problem of one-way structure of backbone network of Faster-RCNN model,semantic and spatial information of this structure for small-scale goods will disappear with the increase of sensitivity field,a commodity detection algorithm(AFF)based on adaptive feature fusion is designed.Firstly,an adaptive multi-scale feature fusion structure(AMFF)is constructed using multi-level feature information to extract feature information from multi-scale targets.Then,the local feature fusion residual module(LFFR)is designed using the characteristics of small convolution kernel to minimize the receptive field and avoid the problem that the network loses the feature information of small-scale targets in the process of multiple down-sampling and convolution,so as to improve the detection accuracy of small-scale targets.Finally,the K-Means clustering algorithm was used to regression the Anchor size suitable for RPC commodity data set,so as to avoid misjudging positive and negative samples and causing the problem of missing detection,so as to improve the overall detection accuracy.In order to solve the problem that AFF detection model is slow,a SESC algorithm based on deep separable convolution is designed.This algorithm uses deep separable convolution-and space separable convolution to construct lightweight classification and regression networks.The compression and extension residual module(DRSE)is proposed to filter out the interference information around the commodity,improve the horizontal and vertical positioning accuracy of the commodity,reduce the network parameters,and greatly improve the detection speed.The spatial channel attention mechanism(DSC)is proposed.The channel attention mechanism integrates weighted local channels into the original channels to form channel features with priority.The spatial attention mechanism integrates weighted local features into the original feature map to enhance the spatial information of the target and improve the localization accuracy.Finally,a new activation function nReLU6 is proposed to reduce the influence of depth-separable convolution on low dimensional features.The accuracy of the SESC detection model mAP50 reached 99.19%,which was improved by 13.38%,2.62%,6.4%,2.5%,1.28%、4.66%、5.94%、3.76%、7.6%and 0.55%compared with the Faster-RCNN,Syn+Render,Retinanet,improved Retinanet,DPNet、YOLOV4-tiny、YOLOV4-lite、E-YOLOV4-lite、SSD and Efficientdet-D3 models,respectively.The SESC detection model reached 0.055s per image,which was 0.201s,0.09s and 0.023s less than the Retinanet,the improved Retinanet and Efficientdet-D3 model,respectively.