Detection of netting knots and main lines in deep-sea cages using an improved U-Net

Addressing the urgent requirements for visual navigation of underwater net-cleaning robots and netting maintenance in deep-sea cage aquaculture, this paper proposes a method for recognizing netting structural features, which is an improved method based on the U-Net model. This method provides reliable visual positioning references and path-tracking foundations for underwater net-cleaning robots by segmenting the key features of netting structure, namely the intersections of mesh wires (Referred to as net knots) and cable lines, while facilitating the detection of damage and fouling on the netting structure. First, underwater netting image data were collected using an underwater net-cleaning robot, and annotations of net knots and cable lines in the images were performed to construct a dataset for netting structural features. Then, using the lightweight semantic segmentation model U-Net as the baseline model, innovations were made by integrating Atrous Spatial Pyramid Pooling (ASPP) and Criss-Cross Attention (CCA) mechanisms. This integration enhanced the model's capability to capture multi-scale contextual information and focus on critical structural regions, thus strengthening the semantic representation of netting features. Additionally, we designed task-specific evaluation metrics tailored to net knots and cable lines, and conducted systematic performance comparisons between U-Net variants with different intermediate-layer channel numbers and the proposed improved model. Experimental results demonstrate that the improved model achieved optimal performance on the constructed netting dataset. Notably, net knots segmentation achieved 0.60 Dice and 0.92 F1, while cable line segmentation attained 0.74 Dice and 0.62 F1, representing significant improvements over baseline configurations. The improved U-Net model integrating ASPP and CCA mechanisms provides an effective solution for precise segmentation of netting knots and cable lines in underwater environments, which also provides a reference for future research on net knot and cable tracking.