Underwater Object Classification in SAS Images Based on a Deformable Residual Network and Transfer Learning

Abstract

To solve the problem of low classification accuracy caused by differences in object types, shapes, and scales in SAS images, an object classification method based on a deformable residual network and transfer learning is proposed. First, a lightweight deformable convolution module DSDCN was designed by adding offsets to a traditional convolution, to adapt to objects with different shapes in SAS images, and the depthwise separable convolution was used to optimize the module. Second, a deformable residual network was designed with the DSDCN, which combined the traditional depth features with deformable features for object representation and improved the robustness of the model. Furthermore, the network was trained by the transfer learning method to save training time and prevent model overfitting. The model was trained and validated on the acquired SAS images. Compared with other existing state-of-the art models, the classification accuracy in this study improved by an average of 6.83% and had an advantage in the amount of computation, which is 108 M. On the deformation dataset, this method improved the accuracy, recall, and F1 scores by an average of 5.3%, 5.6%, and 5.8%, respectively. In the ablation experiments of the DSDCN module, the classification accuracy of the model with the addition of the DSDCN module improved by 5.18%. In addition, the training method of transfer learning also led to an improvement in model classification performance, reflected in the classification accuracy, which increased by 7.4%.