Retail Product Image Recognition Method Based On Deep Learning

With the development of science and technology,more and more scenes are be-coming automated or semi-automated.After“New Retailing" was proposed,auto-mated retail has received much attention in recent years.Traditional identification meth-ods include barcode-based identification and radio frequency-based identification(RFID).However,barcode-based recognition requires manual assistance,which results in low automation.RFID has also not been widely applied due to high cost.It is an important research field in using artificial intelligence and computer vision methods to automati-cally identifiy retail products purchased by customers.It is the focus to identify products with rotatability,multi-view and less training data under non-constrained conditions.This thesis aims at the problem of supermarket product image recognition.The main tasks are as follows:1.For the recognition of single product image,image recognition in real scene is usually different from the ordinary recognition dataset.The images in real scene contain large area of background and there is no labeling location of product in image.The collection of a large amount of image data requires huge resources,which results in little training data.Moreover,the retail product object has uncertain angle and view which is different from general object.To solve these problems,we present a location labeling algorithm to label product images,which only trained by public general object dataset.Images of different angles are generated after using data augmentation method.Product image recognizing model is trained based on transfer learning with these images and location label.The experimental results show that the product image recognition achieves 86.6%top-1 accuracy and 94.34%top-3 accuracy.2.To recognize and locate multiple retail products in image under the condition of we only have single product training images without location labeling.We first modify the bounding-box regression layer of the Faster RCNN,and propose a class-agnostic bounding-box regression layer.It only need trained with public dataset,but does not require be retrained on the target product image dataset when applied to product local-ization.Combined with Grabcut method,a data augmentation method is proposed to generate a large number of multi-product training images for model training.In ad-dition,a re-identification layer is proposed to improve the recognition accuracy of the class-agnostic Faster RCNN.Finally,we achieve 93.8%recall and 96.3%precision.3.Combining the above methods,a retail product image recognition verification system is designed and implemented,and the proposed algorithm is verified and tested.