Abstract:Sonar image classification is crucial in salvage operations and submarine pipeline detection. However, it faces challenges of low resolution, few-shot, and long-tail due to multipath interference and data collection issues. Current methods employ transfer learning, resampling, and adversarial attacks to address these challenges. Nonetheless, knowledge transfer from optical to sonar images is often inefficient due to significant domain differences. Furthermore, the large receptive fields of existing models make it difficult to extract local details from low-resolution sonar images. Additionally, cross-entropy loss excessively suppresses tail class gradients, causing a bias towards head classes. To address these problems, this paper proposes a Hierarchical Transfer Progressive Learning based on the Jigsaw puzzle and Block Convolution (HTPL-JB). First, we introduce a hierarchical pre-training strategy incorporating a source pre-training phase into the transfer learning phase, enhancing the efficiency of transferring knowledge from optical to sonar images. In the fine-tuning phase, we employ a progressive training strategy to progressively extract information at different granular levels, enhancing the model's ability to capture fine details from sonar images. Finally, we introduce a key point sensitive loss (KPSLoss), which uses a larger margin distance and a smaller slope factor for the tail class to enhance accuracy and the separability of key points. Extensive experiments on the NKSID datasets demonstrate that HTPL-JB significantly outperforms the existing methods. Our code will be available at https://github.com/leeAndJim/JBHTPL.

Hierarchical and progressive learning with key point sensitive loss for sonar image classification