Few sample classification method based on attention model

Abstract

The invention relates to a few sample classification method based on attention model, which comprises the steps of training a convolutional neural network image classification model, removing a full connection layer of the convolutional neural network image classification model after the convolutional neural network image classification model is trained, and reserving a convolutional neural network part of the convolutional neural network image classification model; performing basic classification on the to-be-detected image according to the feature vectors of all the sample images to obtain n* m scalar quantities, n being the number of classes of all the sample images, and m being the number of sample images in each class; connecting m sample images of each type into an m * M-dimensionalvector, and inputting a nonlinear mapping function to obtain n * m relative weights; multiplying the obtained n * m scalar quantities by the corresponding n * m relative weights to obtain new n * m scalar quantities, and adding the new m scalar quantities in each class to obtain a final classification result. According to the method, by mining the score of the importance of the intra-class images, the model can pay more attention to important samples, and the intra-class images are balanced.

Description

technical field [0001] The invention relates to a sample classification method. In particular, it concerns an attention model-based few-shot classification method for computer vision. Background technique [0002] The recognition of image objects is one of the most basic and valuable directions in the field of computer vision, and it is the basis of other image processing methods. At present, most image recognition is based on deep learning networks. Although very high classification accuracy can be achieved in many scenarios, these methods require millions of data sets to support network training, and the acquisition cost of samples is extremely high. At the same time, the training and use of the network requires a very long time and a large amount of computing resources. It usually takes several days on the current advanced GPU equipment. After training, it may take several seconds to recognize an image. It is unacceptable especially for some embedded devices. Although i...

AI Technical Summary

Problems solved by technology

[0004] Although with the increase of researchers, image recognition with a small number of samples has made some progress in recent years, most of the work focuses on learning a measure or designing a measure method, resulting in many methods with minimal innovation, and these works pay little attention to samples. The amount of information contained in the picture itself is different, and the important information contained in the same type of picture itself has not been discovered, so although some progress has been made, the effect is not very good