UTVit: A U-Shaped Segmentation Network for Underwater Images Based on TinyVit

This paper presents a lightweight semantic segmentation model for underwater hazardous objects based on the TinyViT backbone network, realizing a real-time semantic segmentation under the computational constraints of unmanned underwater vehicles (UUVs). To efficiently leverage feature maps from various network layers, we devise a network architecture based on U-Net named UTVit, incorporating skip connections and concatenations to integrate feature maps across different levels of the network. Furthermore, we engineered a specialized upsampling network that employs depth-wise convolutions instead of standard convolutions. This approach mitigates the coarse granularity issues often encountered during upsampling to significantly reduce the computational burden. To validate the effectiveness of our proposed model, we conduct extensive experiments on the publicly available USIS10K dataset, comparing our model with state-of-the-art (SOTA) models. The results show that UTVit achieves a slight 2.3% decrease in mIOU (mean Intersection over Union) while significantly reducing the number of parameters by 51.46M, remaining a 12.7% parameter count of the former. In the parameter-mIOU and computation-mIOU trade-off charts, UTVit exhibits superior performance, making it highly suitable for object segmentation tasks in UUV applications. This balance between efficiency and effectiveness underscores the model's potential for practical deployment in resource-constrained underwater environments.