A Seabed Sediment Semantic Segmentation Method Based on Deep Transfer Learning with Self-Attention Mechanism

Abstract

Semantic seabed sediment segmentation is one of the key research contents in the fields of marine resources development and marine engineering. In today’s intelligent and digitalized world, more refined semantic segmentation of seabed sediment can provide the basis for downstream tasks as building digital seabed models. In this paper, a seabed sediment semantic segmentation method based on deep transfer learning with self-attention mechanism is proposed. The overall framework is based on the UNet model. The transfer learning method of parameter sharing is used, while the parameters of backbone pretrained from ImageNet datasets are used as the initial weights of model. According to limited acoustic data of seabed sediment, this paper uses original side-scan sonar strip images to produce dataset. The proposed method is also compared with traditional machine learning methods as SVM (Support Vector Machine) and RF (Random Forest). The self-attention UNet (backbone: VGG16) model performed the best, with mIOU of 92.28%, mPrecision of 94.78%, and mRecall of 96.84% on the test set. The self-attention mechanism improved mIOU by 2.5%. The mIOU of deep learning methods is on average over 20% higher than that of traditional machine learning methods.