Research On The Method Of Commodity Image Detection

Supermarkets play an important role in people's lives.With the ever-expanding large-scale supermarkets and the idea of "unmanned supermarkets",the demand for intelligent supermarket management is also getting higher and higher.Traditional supermarket management relies mainly on manpower,and has the problems of cumbersome management and low efficiency.At present,the artificial intelligence technology is developing rapidly.By collecting the product images in the supermarket surveillance camera and automatically judging and analysing the product attributes through software algorithms,it is helpful to assist the supermarket management.Based on computer vision technology and deep learning method,this thesis studies,designs and implements a set of commodity image detection methods.The article image detection method based on the Faster RCNN detection framework is mainly based on region recommended network regional advice,using convolutional neural network as a feature extractor and classifier to achieve the end-to-end network structure of commodity image detection methods.In this paper,according to the special application scene of commodity image detection,the basic network model is improved for the easy occlusion of commodity images and the inclination of the shooting angle,the introduction of image context information in the regional suggestion technology to improve the accuracy of network classification,the introduction of feature channel weighting technology to enhance feature extraction and the effectiveness of introducing online difficult sample mining technology to improve the accuracy of the classifier.In the experimental phase,improved methods such as image multi-scale and image flipping were added,finally reaching 84.1%mAP in the commodity image detection dataset.The main innovation of this article are as follows.(1)Aiming at the application scenarios of commodity image detection,this paper establishes a special commodity image detection dataset,which contains multiple categories,different lighting,different distances and different numbers of samples,which can effectively characterize the commodity images,Ability to support training and testing of image inspection.(2)In this paper,the fusion of image contextual information,feature channel weighting and on-line hard-to-do sample mining techniques are used to effectively improve the accuracy of commodity image detection methods.(3)In this paper,the deep learning detection method is applied to the detection of commodity images,which enables the computer to automatically obtain the location and category information of the commodities,intelligently assist the supermarket management and effectively improve the management efficiency of the supermarket.