Research On Commodity Detection Algorithm Based On Deep Learning

Self-service supermarket is convenient and efficient,and has gradually become a new shopping mode.Different from the self-service code scanning settlement method of traditional supermarkets,the settlement process of self-service supermarket occurs at the end of shopping action and is faster.Therefore,accurate detection the commodity information in the shopping process in the self-service supermarket environment is crucial to mastering the commodity sales in real time and providing better shopping experience for customers.According to the shopping habit of ordinary people,the acquisition of customer shopping information can be realized through the detection of hand-held commodity.Based on computer vision technology and deep learning theory,this paper studies multi-object commodity detection,single-object background hand-held commodity recognition and multi-object background hand-held commodity detection.The main research contents are as follows:1.In view of the lack of multi-target background hand-held commodity detection dataset,this paper constructs a dataset containing both color images and depth images with the simulated self-service supermarket commodity shelves as the background and 10 kinds of commodities as the detection targets,which is suitable for the multi-target commodity detection and the multi-target background hand-held commodity detection.2.Studied the deep learning target detection structure Faster R-CNN to construct a multi-target commodity detection model.Aiming at the phenomenon that Kinect key point detection method of human body is prone to skeleton disorder when body occlusion occurs,this paper studies the key point detection method of human body based on CPMs model,and obtains the location of key points by obtaining the texture information of images and the spatial information of the confidence map of key points,so as to provide support for the realization of handheld commodity detection model.3.Aiming at the single target background hand-held commodity recognition,this paper studied the segmentation of commodity image based on region growth algorithm to extract convolution and SIFT features respectively and fuse them with heterogeneous features.Proposed a Bagging-SVM ensemble classifier to solve the problem of accidental classification results of a single classifier.Experiments show that the recognition result of ensemble classifier reaches 86.25%,which is 2.09% higher than that of single classifier.4.In view of the multi-target background hand-held object detection,calculate the position information of the product and hand on the two-dimensional image,fuse the depth image,and the corresponding depth is compared to obtain the hand-held object.In order to reduce the irrelevant background input of the detection model,a local amplification algorithm based on hand position was adopted to improve the multi-target background hand-held object detection.The detection accuracy of the improve model reaches 93.26%,which increased by 4.4%.In this paper,researched of commodity detection algorithm based on deep learning,it provides theoretical basis and implementation scheme for the pure visual commodity settlement mode of self-service supermarket,which can effectively improve the commodity settlement efficiency of self-service supermarket.