Continuous learning image classification method and device based on deep learning

Abstract

The invention discloses a continuous learning image classification method and device based on deep learning, and the method comprises the following steps: constructing a task continuous learning model with task specific batch normalization, and enabling parameters of all convolution kernels in a feature extractor of the task continuous learning model to be fixed in all tasks, when each new task is learned, parameters of a corresponding batch normalization layer BN in each convolution kernel and a specific classification head of the task are learned together; incremental training is carried out on the task continuous learning model, and when a new task comes, a specific batch normalization layer and a classification head of the new task are added; after incremental training is completed, a trained task continuous learning model is obtained, image tasks to be classified are input into the trained task continuous learning model, and the classification tasks are completed. According to the method, the problem of disastrous forgetting is effectively solved by utilizing batch normalized BN existing in a task continuous learning model.

Description

technical field [0001] The present invention relates to the technical field of image classification, in particular to a continuous learning image classification method and device based on deep learning. Background technique [0002] Artificial intelligence, especially deep learning, has achieved great success in image classification. However, there is a problem with the model, the catastrophic forgetting problem. That is: after learning new knowledge, the previous training content is almost completely forgotten. In short, such a problem is that in the process of continuous learning, in the learning of each task, the model will be updated with different data, and learning new tasks will greatly reduce the performance of old tasks. Therefore, it is crucial that deep learning models have a continuous learning ability: that is, the ability to continuously learn new tasks without forgetting how to perform previously trained tasks. [0003] In recent years, a variety of methods...

AI Technical Summary

Problems solved by technology

But this extension-based approach inevitably changes the structure of the classifier, especially the feature extractor part, and requires more and more storage space for storing the extended network model

Additionally, determining when and where to add new network components can be extremely challenging

Images