<p>Unmanned underwater vehicles (UUVs) have been employed for the inspection of net damage in aquaculture farms. However, challenges such as dim underwater environments and net folding make it difficult for actual damage to be distinguished from normal folds in fishing nets by UUVs. To address this issue, a visual inspection method for detecting damaged nets under low-light conditions and folding interference using UUVs is proposed in this paper. Potentially damaged areas are first identified by comparing the areas of adjacent mesh openings. This design offers the advantage that false alarms caused by area fluctuations due to non-damage factors, such as overall or regional stretching and compression induced by water currents, are effectively suppressed. To address missed detections due to inadequate underwater lighting and net folding, a YOLOv8 model integrated with an attention mechanism is employed. Sparse connectivity and deformable convolution are utilized by this mechanism. Secondary verification of the initial detection results is performed by the model. Consequently, the damaged sections of aquaculture nets are accurately identified. Initially, underwater net images are preprocessed to extract the net structure and are binarized. Subsequently, each mesh opening is assigned a number, and meshes adjacent to a reference mesh are grouped. The mesh openings within each group are sorted in ascending order based on their area. If the difference between the largest and second-largest areas within a group exceeds the average mesh area of that group, it is classified as damaged. To address missed detections caused by net folds, a feature recognition approach is proposed. Slender net gaps are targeted by this method, which incorporates sparse connectivity, deformable convolutional attention mechanisms, and contextual information mining. Its purpose is to enhance the accuracy with which holes in fishing nets are detected by UUVs. Furthermore, to tackle issues arising from poor underwater illumination, a weakly supervised transformation robustness module is applied. This module is based on covariant regularization and facilitates deeper feature interactions. Consequently, the model's precision in identifying net damages under dim lighting conditions is further improved. Experimental results demonstrate that an accuracy of 92.00%, a recall rate of 82.10%, and a mAP of 87.30% are achieved by the proposed model. Compared with other models, the highest detection accuracy under both low-light conditions and scenarios involving folded nets is exhibited by the proposed model. Thus, an effective solution for UUV-based inspection of underwater net damage in poorly lit environments is provided.</p>