Incremental image classification method and system

Abstract

The invention discloses an incremental image classification method and system. The method comprises the following steps: dividing an obtained training sample set into a plurality of groups; training a Bayesian neural network model through multiple groups of training samples, and classifying images through the trained Bayesian neural network model. In the training process of the Bayesian neural network model, in order to relieve the problem that the classification capability of the model is reduced due to weight importance accumulation, a weight selection mechanism is introduced, and after the weights are screened, the regular strength of part of the weights is diluted in a targeted mode. According to the method, in a scene with fixed capacity, the classification capability of the model for future image samples can be maintained to the maximum extent, meanwhile, the classification precision of old tasks can be protected, and the problem that the classification capability of a regularization-based incremental image classification model cannot be maintained for a long time is effectively relieved.

Description

technical field [0001] The invention relates to the technical field of image classification, in particular to an incremental image classification method and system. Background technique [0002] Deep neural networks have achieved great success in the field of image classification. However, most of today's deep networks recognize and classify images in batch mode, that is, all image samples need to be input into the model as a whole for training and classification. When faced with an incremental scene where new samples may appear at any time, the model parameters will be adjusted to adapt to the new image samples, which will lead to a decline in the ability of the trained model to recognize old images. In order to enable the model to maintain high image classification performance in incremental learning scenarios, it is necessary to overcome or alleviate the catastrophic forgetting problem (catastrophic forgetting) in the incremental learning process. [0003] As early as t...

AI Technical Summary

Problems solved by technology

First of all, most of them judge the importance of a single weight, but in doing so, different weights of the same node will be regularized with different strengths, which will change the information flowing to the node through the weight , resulting in the interference of the features learned by the nodes

Secondly, when the model capacity is fixed and cannot be expanded, the above methods will continuously accumulate the importance of model parameters in order to consolidate the knowledge acquired from old data. Such a method will seriously weaken the learning ability of the model and cause new tasks to fail. Degraded performance due to inability to obtain sufficient learning resources

Images