Research on sonar image few-shot classification based on deep learning

小样本下基于深度学习的声呐图像分类研究

¹ School of Marine Science and Technology, Northwestern Polytechnical University, Xi’an 710072, China
² Key Laboratory of Underwater Information and Control, the 705 Research Institute of CSSC, Xi’an 710077, China

Received: 5 November 2021

Abstract

The underwater environment is complex and diverse, which makes it difficult to evolve traditional methods such as manually extracting features from blurred images. What's more, sonar images are so hard to be obtained that their number is far less than optical images, this case usually is called as few-shot, which leads to over fitting and low recognition accuracy of networks for sonar image classification. Based on the established sonar image data set after image preprocessing, a sonar image few-shot classification method with multi strategy optimization fusion is proposed in this paper. It is an improved label smooth regularization method with category preferences that can optimize the labels of training data and reduce the self-confidence of the network. And then based on the fine-tuning method in migration learning, some parameters of pre-learned models from optical images domain are utilized to help improve the performance in the sonar images domain. Finally, all the above three optimization strategies are combined. The simulation experiments in this study conclude that the optimal recognition accuracy can increase to 96.94%, which proves the multi strategy fusion can effectively suppresses the overfitting phenomenon and accurately classifies sonar images in the case of few-shot.

摘要

水下环境复杂多样, 使得声呐成像模糊难以人工提取特征, 同时声呐图像不易获取, 数量远少于光学图像, 导致了小样本情况下声呐图像分类网络的训练过拟合现象明显, 识别准确率低。基于所建立的声呐图像数据集进行预处理后, 提出一种改进的带有类别偏好的标签平滑正则化方法, 对训练数据的标签进行优化, 减轻网络的自信程度, 并基于迁移学习中微调的方法利用光学图像对网络参数进行预训练和冻结, 融合以上方法构建了一种小样本下的分类网络模型。仿真实验结果表明, 优化后的网络模型取得了最佳分类识别准确率, 有效抑制了过拟合现象, 能够在小样本下实现精确分类声呐图像。