Underwater object segmentation using improved DeepLabv3+

Underwater object segmentation plays a crucial role in marine exploration, ecological monitoring, and robotic vision. However, challenges such as light attenuation, color distortion, and reduced visibility in underwater environments severely impede segmentation accuracy. In this paper, we propose an improved DeepLabv3+–based method for underwater object segmentation that integrates a dynamic multi-scale fusion module, the CBAM attention mechanism, and a deep supervision strategy into the original network to enhance its ability to model objects of varying scales and complex backgrounds in underwater scenes. Our main contributions are as follows: (1) We introduce a dynamic multi-scale fusion module that adaptively merges feature maps with different dilation rates, thereby improving the network’s sensitivity to underwater objects at multiple scales; (2) We incorporate the CBAM attention module into the segmentation network to reinforce feature representations by applying channel-wise and spatial attention; (3) We employ a deep supervision strategy by adding an auxiliary loss to intermediate feature maps, which accelerates training convergence and boosts segmentation performance. Experiments on the SUIM dataset demonstrate that the proposed method achieves an 88.46% accuracy in detail reproduction and boundary precision, effectively overcoming the limitations of traditional underwater segmentation approaches.