Detecting underwater structural defects is vital for hydraulic engineering safety. Diverse patterns of underwater structural defects, i.e., the morphology and scale characteristics, pose difficulties on feature representability during detection. Any single feature morphology is insufficient to fully characterize diverse types of underwater defect patterns. This paper proposes a novel hybrid neural network to enhance feature representation of underwater structural multi-defects, which in turn improves the accuracy and adaptability of underwater detection. Three types of convolution operations are combined to build Hybrid Aggregation Network (HanNet), enhancing the morphological representation for diverse defects. Considering the scale difference of diverse defects, the Multi-Scale Shared Feature Pyramid (MSFP) is proposed, facilitating adaptive representation for diverse sizes of structural defects. The defect detection module leverages an Adaptive Spatial-Aware Attention (ASAA) at the backend, enabling selective enhancement of salient defect features. For model training and evaluation, we, for the first time, build an underwater structural multi-defects sonar image dataset containing a wide range of typical defect types. Experimental results demonstrate that the proposed model outperforms state-of-the-art methods, significantly improving defect detection accuracy, and provides an effective solution for detecting diverse structural defects in complex underwater environments.

Small object detection remains a significant challenge in computer vision, primarily attributed to insufficient feature information and low resolution. This paper introduces SAMYOLO, a novel framework that integrates super-resolution (SR) technology with a mamba-based feature enhancement strategy to improve small object detection. SAMYOLO incorporates a dual-branch design to process both RGB and super-resolved images, improving detail preservation and feature representation. A dedicated feature mapping (FM) module, built upon the State Space Model (Mamba), is designed to reinforce critical features and enable effective cross-modal fusion at multiple scales. Furthermore, an optimized loss function is introduced to enhance detection accuracy and training stability. Experimental results on the SIMD and VEDAI datasets demonstrate SAMYOLO's superior performance, achieving significant improvements in mAP50 over state-of-the-art models. The framework excels in dense small object scenarios, highlighting its practical potential for applications such as remote sensing and aerial imagery analysis.

Image super-resolution (SR) and disparity estimation are closely related in stereo images, and exploiting the features within disparity maps effectively enhances the performance of SR. In this paper, we proposed a stereo image super-resolution reconstruction network based on disparity estimation and domain diffusion model. The network combines disparity estimation and diffusion models to perform feature summation and construct a domain space, facilitating interaction between forward and backward diffusion memory units. To further enhance the quality of the reconstructed images, we introduce a different memory unit using the combinatorial concept of long short-term memory. Furthermore, disparity loss and structural similarity are incorporated into the loss function, aiming to yield more accurate results. Experiments demonstrate that our method outperforms existing approaches and achieves a 23% reduction in the number of parameters. Finally, the proposed method is applied to underwater stereo images. The results showcase excellent prospects for application.