Research On Supermarket Product Detection And Recognition Algorithm Based On Deep Learning

At present,the settlement of supermarket commodities is mainly realized through barcode or RFID tag technology.The barcodes need to be scanned manually,which leads to low efficiency.RFID tags are expensive and cannot be reused.Therefore,how to use artificial intelligence technology to realize fast automatic detection and recognition of supermarket products.Automatic detection and identification to realize automatic settlement has become a research hotspot in recent years.On this basis,a supermarket product detection and recognition algorithm based on deep learning is proposed to reduce the cost of product settlement and improve efficiency.In view of the wide variety of supermarket products and their local characteristics,this article starts with the application scenario,using the left,top and right cameras to collect images of different sides of the product on the conveyor belt,and references the SIFT image stitching algorithm based on phase correlation and texture classification.Commodity images are spliced,and the sensitive features and overall characteristics of the commodities are detected through image detection algorithms.Finally,the product identification algorithm based on multi-feature fusion of random forest is used to realize the rapid identification of commodities.The main research content of this article includes the following three aspects.(1)Aiming at the shortcomings of low accuracy and slow speed of Retinanet network model detection,an image detection algorithm based on improved Retinanet is proposed.The algorithm first uses information interaction fusion technology and improved feature pyramid to extract image texture features and high-frequency abstract features,and then uses SSH detection network to replace the classification regression sub-network in the original network model to expand the image detection area,and finally uses weighted loss to perform Regression to improve the efficiency of image detection.In the RPC data set,compared with the original Retinanet network model,the detection accuracy of the algorithm in this paper is increased by3.90% to 96.69%;the average detection time of a single image is reduced from 0.201 s to 0.145 s,and the detection rate is increased by 27.86%.(2)Aiming at the high-confidence features such as barcodes and text sequences in the outer packaging of commodities,a barcode detection algorithm based on T-Densenet and a WhCTPN text sequence detection algorithm are proposed.Based on the T-Densenet barcode detection algorithm,the Densenet network layer is added and the anchor value is re-adjusted to increase the barcode coverage area,thereby accurately and quickly detecting the barcode characteristics of the product.The Wh-CTPN text sequence detection algorithm redefines the Side-refinement formula in the CTPN algorithm to achieve accurate positioning of text sequence information,and then extracts important text information such as product trademarks,specifications,and text descriptions.(3)Aiming at the diversity of commodity features,a multi-feature fusion commodity recognition algorithm based on random forest is proposed.The algorithm uses Zbar barcode recognition algorithm,Simi-Dense Net-CTC text sequence recognition algorithm and SCTDNet recognition algorithm to recognize the barcode,text sequence and overall characteristics of the product,and enters the result into the random forest product feature classifier for classification and recognition.The budget score of each input feature is output,and the optimal recognition result is output.Finally,information is matched through the product image database to achieve efficient product detection and recognition.In the self-made product image data set,compared with the original Retinanet algorithm and the improved Retinanet algorithm,the product recognition accuracy rate of the algorithm in this paper is increased by 6.11% and 2.48% to97.81%,respectively;the average detection time is shortened to 0.18 s,and the detection rate is increased respectively.33.33% and 18.18%,which can better realize the rapid identification of commodities.Experiments have proved that this article has achieved good results in both the public data set and the self-made commodity image data set,verifying the effectiveness and feasibility of the algorithm.