A Novel Underwater Blurred Target Detection Algorithm Based on RT ‐ DETR

Underwater target detection is crucial for monitoring marine resources and assessing their ecological health. However, the underwater environment is complex and variable, and factors such as light attenuation, scattering, and turbidity often cause optical images to be blurred and target details to be unclear, seriously affecting detection accuracy. Although deep learning‐based methods have shown promise in the field of target detection, challenges remain in balancing real‐time performance with high‐precision detection of blurred targets. In response to the above situation, a novel algorithm is presented for underwater blurred target detection, designed to address the challenge of low detection accuracy resulting from indistinct optical image details in complex underwater environments. The proposed algorithm leverages the Real‐Time Detection Transformer (RT‐DETR) architecture. First, a lightweight feature extraction module, termed Faster‐Rep (FARP), is developed to effectively reduce the model's parameter count while simultaneously enhancing the backbone network's ability to extract salient features from blurred targets. Second, an efficient additive attention module, called AIFI‐Efficient Additive Attention (AIFI‐EAA), is utilized in the coding phase, which enhances the model's global modeling capability while significantly reducing computational redundancy. Atlast, the Dynamic Cross‐Scale Feature Fusion (DyCCFM) module enables dynamic fusion of feature information, thereby preserving critical characteristics of blurred targets and preventing information loss. The proposed algorithm demonstrates excellent detection performance on the URPC2020 dataset, where the mean Average Precision (mAP) is improved by 1.5% and the number of parameters is reduced by 14.5%, which also significantly improves the ability to detect ambiguous targets in intricate underwater environments.