Research On One-shot Retail Product Recognition

With the rapid development of mobile Internet,vision-based retail product recognition plays a key role in automated checkout,marketing promotion and intelligent after-sales service.Most of the current retail product recognition algorithms rely on convolutional neural networks and optimize the model by big data.However,due to the wide variety of commodities,it takes a lot of manpower and resources to collect enough training samples for each product.For this reason,this paper proposes an algorithm of one-shot retail product recognition.In this paper,we divide the problem of one-shot retail product recognition into two sub-problems.One is to retrieve similar images of the query from a large retail product dataset.The dificulty of this problem lies in how to map the product image to vector under the condition of single sample.The other is to make accurate recognition based on geometric constraints.The difficulty lies in how to ensure the high robustness of local features and improve the speed as much as possible.For the two sub-problems,the main work of this paper is as follows:(1)This paper proposes an image embedding algorithm based on capsule neural network to realize the mapping from retail product image to vector.Compared with the convolutional neural network,the robustness of the model to the change of viewpoint is significantly improved.In order to further improve the performance of the model,this paper also uses spatial transformation network to align the features and proposes a new reconstruction module,which achieved better results while reducing the number of parameters.Finally,this paper proposes a ladder training method,which can effectively compensate the performance loss caused by the difference in training quantity.The test results show that under the condition of single sample,compared with the convolutional neural network,the image embedding algorithm proposed in this paper improves by 9.1 percentage points,reaching 68.2.(2)For the problem of accurate recognition of retail product image,this paper designs a deep binary local feature which can be used as the basis for solving geometric verification.In this paper,deep convolutional network is used to extract local features of images,which achieves good robustness for illumination,perspective and other factors.At the same time,through the special hidden layer design,to achieve the binary representation of features,in terms of speed,compared with other depth local features,has an obvious advantage.Experiments show that the proposed local feature extraction algorithm has the same accuracy as the current algorithm,and has an obvious advantage in speed.It can reach 22FPS for the images of 640*480.(3)Based on the above two algorithms,this paper designs and implements a retail product recognition system.The system can recognize other images of the product by simply uploading several product template images.At the same time,the system can also use the image data uploaded by users to optimize the algorithm itself.