Research On Commodity Recognition Technology Based On Deep Learning

With the improvement of people's living standards,a large number of supermarkets and department stores can be seen everywhere in daily life,which greatly meet the material level of people's life.The development of Internet technology has resulted in a variety of sales patterns,and the automatic identification technology of commodities is one of the key technologies for the efficient circulation of commodities.In the field of "new retail",people are constantly trying to reduce the increasing labor costs through scientific and technological means,so the prospect of unmanned retail stores is very good.One of the key technologies of no-man's shop is the automatic identification technology of goods.Accurate and efficient automatic identification technology of goods can not only reduce labor costs,but also increase the satisfaction of customers' shopping experience.Therefore,it is of great significance to design an efficient and accurate method of commodity identification.Through the study of image detection,image classification and loss function in deep learning,a method of automatic commodity recognition with less training data,low cost and high accuracy is designed.The main research contents include:Firstly,a data set of 26 kinds of commodities is constructed,and data augmentation is carried out according to the problems that may be encountered in actual production.Through the analysis of data set images and actual usage scenarios,the conclusion is drawn that there are many commodity angles and the aspect ratio is complex and changeable in the image,and the image detection method is improved according to the conclusion.The data collected in various scenarios are used for testing.Results The robustness of image detection method in commodity detection was verified.Secondly,through image detection algorithm,200 images are obtained by program discrimination and manual selection,which are randomly divided into training set and test set according to 1:2.Through data augmentation,transfer learning and optimization techniques,the classification results of the model are continuously optimized,and the loss function is used to increase the classification ability of the model.The instruction set based on SMID technology is used to accelerate the reasoning of the network model,so that the reasoning of the model on the CPU is faster.The experimental results show that the image classification algorithm can accurately identify a variety of commodities,and the improved transfer learning and loss function can effectively improve the classification ability of the model when the training samples are small.Finally,the combination of commodity detection and commodity identification identified by the commodity identification system is used to test the recognition speed and accuracy of the commodity identification system.The results show that the time required to test an image is 0.13 s,and the accuracy rate is 99.86%.The performance in accuracy and speed satisfies the requirements in practical use.