Research On Object Detection Algorithm Based On Convolutional Neural Network

As a hot issue in the field of computer vision,object detection has achieved good research results in many applications in recent years.Traditional supermarket management relies mainly on labor,using bar code technology and radio frequency identification technology to identify commodities,with low automation and high cost.Under the development of automated new retail,the use of computer vision technology in the identification of supermarket commodities helps to achieve intelligent management,and has broad application prospects and value.At present,there is no available supermarket commodity dataset for research use.To solve this problem,this thesis collects eleven different commodities in complex scenes of supermarkets and makes a relatively complete supermarket commodity dataset by data enhancement algorithm and manual labeling.In addition,in order to better explore the problems in commodity identification,this thesis first implements the traditional commodity detection algorithm.The segmentation of the commodity area and the identification of the commodity target are two important steps in completing the commodity detection.Due to the weak adaptability of the manual features and the poor robustness,resulting in accumulation of errors and affecting the final recognition accuracy.Object detection needs to weigh the detection accuracy and detection speed.According to the research in this thesis,it is find that YOLO(You Only Look Once)algorithm based on regression has good real-time performance and high detection accuracy.As a simple network,tiny-YOLO is more suitable for the training of small commodity dataset to avoid over-fitting.Therefore,tiny-YOLO network is used for commodity detection.And after various improvements and optimizations,CD-YOLO(Commodity Detection based on improved tiny-YOLO)adapted to the characteristics of commodities is implemented.Firstly,K-means clustering is used to analyze the commodity dataset,and the multi-scale bounding box design is optimized to improve the positioning performance.Then,based on the multi-feature fusion algorithm,the recognition ability of small and unclear objects is improved.Finally,this thesis analyzes the loss function of the network,reasonably sets the proportion of loss weights of each part.Therefore,the network convergence speed is improved and the detection performance is improved.Compared with the traditional detection algorithm,it is shown that the commodity detection algorithm based on regressive convolutional neural network has higher recognition accuracy and good real-time performance.Due to the small number and the lack of clarity of some commodities,it is difficult to be identified,resulting a decrease in detection accuracy.In response to this problem,this thesis proposes an end-to-end commodity detection network FRR-Net(Fusion of Res Net and Regression Network)based on the principle of Res Net(Residual Network).Firstly,through the cascade of multiple bottleneck residual blocks,the feature extraction basic network is built to obtain high-level abstract features of images,and avoid the degradation problem of deep network.Secondly,in order to construct the identity map of the residual block and preserve the complete information of the fine-grained features,the pixel sampling algorithm is proposed to improve the dimension matching method of feature maps.Finally,the object recognition and positioning are completed by combining Softmax classifier and multi-scale bounding box strategy.In the process of training,the batch normalization strategy is introduced to solve the problem of network training divergence and instability.In the process of detection,a non-maximum suppression algorithm is adopted to select the optimal boundary position and improve the Recall value of the network.The detection experiment of commodity images under different conditions shows that the FRR-Net algorithm can better adapt to the complex and diverse background environment,and the average recognition accuracy of various commodities reaches 98.57%,which is robust.