Generalized zero sample image classification method based on recognizable false feature synthesis

Abstract

The invention provides a generalized zero sample image classification method based on recognizable false feature synthesis. The method comprises the following steps: constructing an end-to-end neural network model; pre-training the model by using the visible images, so that the distance between the similar potential features in the potential space and the semantic attributes of the similar potential features is minimum, and the visible distinguishable potential features are obtained; for each unseen class, semantic attributes of the visible class meeting the similarity judgment requirement are selected to construct an attribute transformation matrix, and the attribute transformation matrix is used for optimizing a non-negative composite vector; combining the potential features of the selected visible classes and the semantic attributes of the unseen classes by using the non-negative synthesis vectors to synthesize pseudo features of the unseen classes; filtering the synthesized unseen class false features and removing outliers in the false features to obtain distinguishable false features; and training the whole network by using the distinguishable false features and the visible images. According to the invention, high-precision classification can be carried out on the visible and unvisible images at the same time.

Description

technical field [0001] The invention relates to the technical field of image classification and generalized zero-sample learning, in particular to a generalized zero-sample image classification method based on synthesis of distinguishable pseudo-features. Background technique [0002] At present, the neural network has achieved certain results in the field of image classification, but it needs to be trained with a large amount of image data to achieve accurate recognition. It is difficult for neural networks to identify unseen classes that are not in the training samples. Zero-shot learning (ZSL) is a popular research method in the field of image classification of unseen classes. It constructs a neural network model, uses the existing categories in the training set as visible classes, trains the neural network, and solves the classification problem of unseen classes. The test phase only classifies images of unseen classes. For an excellent classification model, it is neces...

AI Technical Summary

Problems solved by technology

[0007] In view of this, the present invention provides a generalized zero-sample image classification method based on the synthesis of discriminable pseudo-features, which solves the poor recognition ability of unseen classes, the domain drift between visible classes and unseen classes, and the composition of unseen classes. See the deviation between the pseudo-features of the class and the real features to improve the classification accuracy

Images