An underwater image segmentation model for complex scenes in aquaculture using vision Transformer

Abstract

Underwater image processing holds significant importance for aquaculture. Traditional segmentation models have primarily been trained and tested in terrestrial environments. However, substantial differences exist between underwater and terrestrial environments, particularly in lighting conditions and color distribution, and other underwater physical phenomena. The unique underwater conditions render traditional segmentation models unsuitable for direct application, and obtaining ample data becomes challenging. Consequently, underwater image segmentation techniques encounter issues of diminished segmentation performance in practical applications. Therefore, this paper constructs an underwater image segmentation model for complex scenes in aquaculture. This paper first proposes a data preprocessing method based on Mosaic data enhancement technology to process limited datasets and increase their richness and diversity. Second, a residual feature enhancement module is constructed to capture global context information and enhance local feature extraction so local and global dependencies are effectively combined. Finally, a multiscale feature residual fusion module is proposed, which uses residual convolution to enhance the efficient fusion of low- and high-level semantic features and solves the semantic gap between the encoder and decoder. This paper conducts extensive experiments on the three datasets FishData, the underwater images segmentation dataset, and the large-scale fish dataset to verify the effectiveness of the underwater image segmentation model proposed in this paper. Compared with the existing classical network model, the $MIoU$, $Acc$, $CPA$ and $F1-score$ of the proposed model reached 92.20%, 97.85%, 92.47% and 92.9%, respectively. Experimental results show that the proposed model performs better in underwater image segmentation tasks.