Research On Commodity Detection And Recognition Algorithm Based On Deep Neural Network

With the rapid development of network technology,unmanned supermarkets have been widely concerned as a new retail model.Unmanned supermarkets are often implemented by placing RFID devices on goods,and the cost is high.In recent years,the development of deep neural networks has been rapid,and the rapid change of target detection algorithms has made it possible to make unmanned supermarket solutions based on computer vision possible.Therefore,this paper studies the commodity detection and recognition algorithms of convenience stores,based on the deep neural network.This paper first studies the performance of the target detection algorithm on the generalized data set.Deep neural networks have the advantage of strong feature expression,and have outstanding performance in detection and recognition tasks.Therefore,existing target detection algorithms are mostly based on deep neural networks.commodity recognition takes accuracy as the first indicator,so this paper focuses on region-based algorithms.The Faster R-CNN model is a representative of the region-based detection algorithm.After in-depth study of the model,this paper improves the problem,the feature extraction network expression feature scale is not rich enough in the Faster R-CNN model,by using multi-scale feature fusion of the Faster R-CNN model,and gets better accuracy on the VOC2007 dataset.Based on the research of generalized datasets,this paper proposes a commodity detection algorithm based on deep multi-scale feature fusion for the convenience store commodity datasets,which has characteristics like large differences in product size,obvious occlusion and more slender commodities.Based on Faster R-CNN,the algorithm transforms the feature extraction network into a deep multi-scale feature fusion network to improve the expression ability of the algorithm.Combined with the improved anchor frame selection scheme,the algorithm can greatly improve the detection capability of slender commodity.And through the migration learning and data augmentation,the performance of the algorithm on the commodity data set is further improved.The effectiveness of the proposed algorithm and model is verified by experimental comparison.The detection and recognition of overlapping commodities,slender products and small-scale commodities are better than the original model.The algorithm proposed in this paper has a detection and recognition accuracy rate of 91.2% on the self-built convenience store commodity dataset,and the detection speed can meet the requirements of real-time detection.