Zero sample target classification method based on pseudo sample feature synthesis

Abstract

The invention provides a zero sample target classification method based on pseudo sample feature synthesis. The method comprises the following steps of obtaining the samples and the annotation information of other categories similar to a category to be identified; secondly, obtaining the semantic descriptions of a visible class and a non-visible class by means of network capture and the like, andconverting the class description information into the semantic vectors through a natural semantic processing model; calculating a similarity score between each unseen class and each visible class; constructing a convolutional neural network classification model, wherein the model is divided into a feature extraction part and a classification part; for each unseen class, screening N visible classeswith highest scores according to the similarity scores, randomly selecting the samples, and inputting the samples into a feature extraction network to obtain the feature vectors; combining the feature vectors of the N visible classes according to the similarity score to serve as the feature vectors of the unseen classes; training the classification network by using the unseen feature vectors, sothat the samples of the category can be accurately identified under the condition that there is no available training samples of a certain category of to-be-identified targets.

Description

technical field [0001] The invention relates to a zero-sample target classification method in mark recognition, in particular to a zero-sample target classification method based on pseudo-sample feature synthesis. Background technique [0002] Supervised learning classification methods have achieved great success in all walks of life. Supervised learning learns classification functions through a large amount of labeled training data. However, for a specific category, collecting a large amount of data and annotating it is very time-consuming and expensive. Inefficiency, and even collecting a small amount of data in some fields is very difficult. Therefore, the research on the zero-sample target recognition method for the application of the target to be recognized with very few samples or even no available training samples has important application value. [0003] Different from the traditional supervised learning method, the purpose of the zero-sample target recognition metho...

AI Technical Summary

Problems solved by technology

[0007] This method mainly involves converting the zero-sample target classification problem into a traditional supervised learning problem by performing pseudo-feature synthesis on unseen classes, so as to effectively overcome the low accuracy of the traditional zero-sample target recognition method.

Images