Multi-scale feature map fusion encoding for underwater object segmentation

Underwater object segmentation presents significant challenges due to the degradation of image quality and the complexity of underwater environments. In recent years, deep learning has provided an effective approach for object segmentation. However, DeepLabV3+, as a classical model for general scenes, shows limitations in achieving accurate and real-time segmentation in complex underwater conditions. To address this issue, we propose a DeepLab-FusionNet, an extended version of DeepLabV3+, specifically designed for underwater object segmentation. The model utilizes a multi-resolution parallel branch structure to extract multi-scale information and employs an improved inverted residual structure as the basic feature extraction module in the encoding network. Structural reparameterization technique is introduced to optimize inference speed and memory access costs during the inference stage. Additionally, a module for linking deep and shallow level information is constructed to reduce the loss of detail and spatial information during downsampling and convolution. Evaluation on the SUIM dataset shows a 3.3% increase in mean Intersection over Union (mIoU) and a speed improvement of 34 frames per second (FPS) compared to the baseline model DeepLabV3+. Further comparisons with other classic lightweight models and Transformer-based models on the UIIS and TrashCan datasets demonstrate that our model achieves good accuracy and balanced computational efficiency in challenging underwater environments. Although there is room for improvement due to overfitting and fixed convolution kernel limitations, future integration with Transformer methods is planned. Our model offers an effective solution for real-time target segmentation for underwater robots, with broad applications in human exploration and development of marine resources. Our codes are available at: https://github.com/sunmer1rain/deeplabv_fusionnet