A Method Of Supermarket Commodity Identification Based On Deep Learning

In the supermarket scenario,merchants and consumers both need real-time information about the products on the shelves.In practice,the information of these products is obtained manually.However,due to the huge number of supermarket products,manually obtaining products information is costly and inefficient.Thus,the vision-based products recognition method has important research significance and commercial value.There are mainly three parts in the supermarket products identification method: the detection of products target,the feature extraction of products target and the recognition of products target.The main steps of commonly used products target recognition methods are as follows: First,the image segmentation technology is used to obtain the target,and then the manual features of the products are extracted and matched with the products in the standard products template library,so as to obtain the recognition results of the products target.However,there are many kinds of products in the supermarket,and the appearance characteristics of the products is ever-changing.Therefore,it is difficult to segment the unknown products target from the images by the common image segmentation techniques.In addition,the manually extracted features have difficulties to express the characteristics of massive products adaptively,which makes these methods unpractical in a real-world supermarket.In order to solve the problems mentioned above,this thesis applies the deep learning algorithm to supermarket product target recognition methods,via detecting the target area of the products through deep neural network and adaptively extracting product features through the convolution layer in the network.The main work and contributions of this thesis are as follows: 1.This thesis applies the deep learning technology to detect and classify the supermarket shelf products.First,to obtain the data set used to train the neural network,we collect the images of shelf products in several supermarkets,label all the product areas contained in the images,and classify the products into preliminary categories.Then we construct an end-to-end product recognition model based on the convolutional neural network and train it on the obtained data set.Then the trained neural network is used to detect the product target area and category in the shelf images.By utilizing the preliminary categories,the speed of product recognition can be improved and the similarities of different types of products can be avoided.The experimental results show that the recognition method based on deep learning can accurately detect all contained product targets in the shelf images and accurately classify the products.2.This thesis implements the feature extraction on products by using convolutional neural network and recognize the products by encoding the products feature by clustering method.Firstly,the output of the last convolutional layer in the monitoring network is expressed as the feature of the product.Secondly,as the size of the target product is different,the size of the obtained product features is also different.To facilitate the recognition operation in next step,we use the Fisher Vector to encode the feature of the target product and convert all the detected product features into descriptors of the same dimension,which can help to supplement the feature information of the product.Then we calculate the similarity between the descriptors obtained after coding and the product descriptors in the standard product model library of the same category,and select the model with the highest similarity measure as the recognition result of the product.We also build a small supermarket product model library and test it in a real-world supermarket scenario.The results show that the proposed product recognition method based on deep learning has high recognition accuracy and is feasible under real scenarios.