An Improved UNet Algorithm Based on Multiscale Features and Attention Modules for Underwater Salient Multi-Target Detection

Underwater salient object detection (USOD) has attracted increasing research attention in underwater exploration for its applications in various underwater vision tasks. However, factors such as low contrast, poor quality, and significant changes in the object scale of underwater images often degrade the performance of USOD, making it challenging to meet practical demands. Therefore, this paper proposes an improved UNet model incorporating multi-scale features and attention modules to enhance its feature extraction capability. To validate our model, we selected the USOD10K dataset, known for its high diversity, complexity, and scalability. The results show that the improved UNet model achieved a mIoU of 77.23%, representing a 6.53% increase over the standard UNet's 70.70%. It is expected that the application of this model helps to enhance detection accuracy. The Gradient-weighted Class Activation Mapping (Grad-CAM) approach enhances the understanding of the decision-making process of the improved UNet model, which focuses on more information at different scales than the base UNet model.