Small sample image increment classification method and device based on embedding enhancement and self-adaption

Abstract

The invention provides a small sample image increment classification method based on embedding enhancement and self-adaption, and the method comprises the steps: obtaining an image increment classification system which is used for carrying out the classification task of a to-be-classified image; to-be-classified images are obtained and uploaded to a system for recognition, when system recognition fails, a small number of images of the category are obtained to serve as training samples, the training samples are calculated through a feature pre-training module to obtain a target prototype, and the target prototype and / or an original prototype are / is adaptively adjusted through a mixed relation mapping module to update all prototypes in the system; classifying and identifying the to-be-classified image of the category; and when the system identification succeeds, the to-be-classified image is classified and identified through the feature pre-training module, the hybrid relation mapping module and the classifier, and a classification result is output. The method is used for enhancing the expandability of a classifier, introducing a hybrid relation mapping mechanism and optimizing prototype representation of a sample, so that a system is gradually applicable to identification of all visible images.

Description

technical field [0001] The invention relates to the technical field of automatic identification, in particular to a method and device for incremental classification of small sample images based on embedding enhancement and adaptation. Background technique [0002] Thanks to the availability of massive labeled datasets, deep learning techniques have achieved remarkable success in many computer vision tasks. Manually labeling data is an expensive and time-consuming process, and there are many types of images, and it is almost impossible to completely label all possible image types at one time. Therefore, most current classification algorithms are developed in a closed and static environment. , developed for one or several specific categories. However, the actual scene is usually dynamic, open, and non-stationary. With the continuous emergence of new categories of samples, the classification algorithm requires a large number of new categories of labeled data and old data to re...

AI Technical Summary

Problems solved by technology

There are mainly two ways in current research: one is to fine-tune the embedding model and unify the classifier with a small amount of new sample data, but fine-tuning the network in a new session will cause the knowledge of the old category to be forgotten; the other is to use the embedding The learning of the representation and the classifier is decoupled, and only the classifier is updated when learning a new task, but the frozen embedding indicates that the network lacks adaptability to the feature embedding of subsequent small-sample incremental tasks, which is not conducive to adaptive learning of new samples

[0006] Therefore, in the existing technology, most of the research work has not yet proposed an effective method to alleviate the problem of dynamic development of real scenes, which limits the development of deep learning technology in related fields.

Images