Research On Image Recognition Method Of Unmanned Supermarket Products Based On Deep Learning

In recent years,along with the rapid development of convenience stores,the concept of "no one supermarket" has also come into being,and the most important thing is the identification of the product image.Choosing a simple and fast commodity identification method,not only can reduce the labor cost,but also can greatly accelerate the consumer shopping settlement time,and then expand the sales volume,from both sides to bring business profits.Therefore,designing an accurate and fast product identification method has an important significance for the unmanned supermarket.By analyzing and comparing the current domestic and foreign many kinds of commodity recognition technology,finally chose the image recognition technology based on deep learning which is more excellent in recognition speed and accuracy.In the actual settlement,the target commodities will be placed in different positions in the settlement area,the deep learning-based target detection method is selected for commodity image recognition,which can prioritize the positioning of the commodities before identification,and can effectively solve the complexity problem of traditional image recognition that requires pre-processing of the image before identifying the background and the target object itself.After the comparison,YOLOV3 has obvious advantages over other algorithms in terms of speed and accuracy among various target detection methods,so YOLOV3 is finally chosen as the research and improvement algorithm.To address the problem of lack of unmanned supermarket product images,the self-built product image training and test sample set was selected.According to the characteristics of different types of products,different shooting methods were used,and the shooting color palette and shooting angle were changed to increase the complexity of the dataset,and finally more than 5000 images of products were collected,and Open CV was used to rotate and scale the images to expand the dataset samples,and finally the product images were marked manually to complete the establishment of the dataset.In order to solve this problem,we choose to combine and improve the method of a priori frame in SSD,and perform the secondary positioning identification of products by doing a priori frame processing on 8 areas around the target center of the product image.This increases the probability of recognizing other commodities in the center point area,thus improving the accuracy of commodity recognition,and at the same time,replacing the original network convolution with void convolution,which expands the receptive field and improves the recognition of small commodities without affecting the resolution.Finally,on the basis of the previous dataset,more than 5000 additional images containing multi-commodity occlusion and small-commodity aggregation were taken to supplement the dataset,and the improved network was retrained again.The final recognition accuracy on the test set reached 98%,while the recognition rate for the products containing masking was also significantly improved compared to the previous set,reaching the level of practical application.