Research on a class-incremental learning method based on sonar images

基于声呐图像的类别增量学习方法研究

School of Marine Science and Technology, Northwestern Polytechnical University, Xi'an 710072, China

Received: 29 June 2022

Abstract

Due to the low resolution and the small number of samples of sonar images, the existing class incremental learning networks have a serious problem of catastrophic forgetting of historical task targets, resulting in a low average recognition rate of all task targets. Based on the framework model of generated replay, an improved class incremental learning network is proposed in this paper, and a new deep convolution generative adversarial network is designed and built to replace the variational autoencoder as the reconstruction model of generated replay incremental network to improve the effect of image reconstruction; a new convolution neural network is constructed to replace the multi-layer perception as the recognition network of generated replay incremental network to improve the performance of image classification and recognition. The results show that the improved generated replay incremental network alleviates the problem of catastrophic forgetting of historical task targets, and the average recognition rate for all task targets is significantly improved.

摘要

由于声呐图像分辨率低、样本数少, 现有的类别增量学习网络对历史任务目标出现了严重的灾难性遗忘问题, 导致所有任务目标的平均识别率降低。基于生成重放的框架模式, 提出了一种改进的类别增量学习网络, 设计搭建新的深层卷积生成对抗网络取代变分自编码器, 作为生成重放增量网络的重构模型, 提升图像的重构效果; 构建新的卷积神经网络取代多层感知机, 作为生成重放增量网络的识别网络, 提升图像的分类识别性能。结果表明, 改进的生成重放增量网络缓解了历史任务目标的灾难性遗忘问题, 显著提高所有任务目标的平均识别率显著提高。