Image classification method of class incremental learning based on self-holding representation extension

Abstract

The invention discloses an image classification method based on class incremental learning of self-holding representation extension. The method comprises the following steps: 1, optimizing by utilizing an initial picture set to obtain a basic image classification network; 2, based on a residual adaptive device, carrying out structure expansion on the initial feature extraction part; 3, inputting a resampled initial class prototype and an incremental class sample, and calculating a classification loss function of the expanded network; and 4, inputting the incremental sample into a feature extraction part before expansion, and calculating a distillation loss function by using an Euclidean distance. And 5, updating the network according to the loss function, and introducing a structure re-parameterization technology to recover the network structure to obtain an incremental classification network. According to the method, the problems that the parameter quantity increases too fast and depends on an extra memory in the process that the image classification network carries out structure amplification based on user data can be solved, so that the demand on network storage is reduced while the image increment classification capability is improved, and the possibility that user privacy is leaked when the image increment classification network is trained is avoided.

Description

technical field [0001] The invention belongs to the field of class incremental learning, in particular to an image classification method of class incremental learning based on self-preserving representation expansion. Background technique [0002] In recent years, research attention has increasingly turned to other aspects of learning due to the tremendous progress deep neural networks have made under fully supervised conditions. An important research aspect is the ability to continuously learn new tasks as the input stream is updated, which occurs frequently in practical applications. Quasi-incremental learning is a very challenging task in continuous learning, which has attracted the attention of many scholars. This task aims to accurately identify new categories without forgetting the learned knowledge of the old ones. [0003] For this scenario, it is particularly time-consuming and labor-intensive to jointly train new and old samples at each stage. In addition, the o...

AI Technical Summary

Problems solved by technology

Although this strategy can fully maintain the performance of the old class and show good performance, during the training process, the overall parameters of the network increase sharply and linearly with the staged training

However, in the actual application scenario of the image classification network, the user's new image data will explode over time, which will lead to the complete collapse of the existing model capacity based on dynamic structure amplification.

Images