EL-DETR: An enhanced localization DETR with spatial attention feature pyramid for high-precision underwater object detection

Global aquaculture is a cornerstone of food security, and precise monitoring of farmed aquatic organisms is critical for advancing smart aquaculture systems. However, existing computer vision models struggle to accurately localize and detect target organisms in underwater farming environments. This limitation is primarily driven by image degradation resulting from turbidity, uneven illumination, and low contrast. It is further compounded by the natural camouflage of aquatic organisms, which blurs boundaries and causes severe foreground-background feature confusion. To address these challenges, this study proposes a multi-module collaborative detection framework named Enhanced Localization DETR (EL-DETR) optimized for aquaculture underwater scenes. The core of the framework is the Spatial Attention-based Feature Pyramid Network (SA-FPN), which integrates two synergistic modules: the Efficient Spatial Localization Attention (ESLA) and the Residual Channel Semantic Refinement (RCSR) Block, aimed at improving detection performance for multi-scale objects in noisy underwater environments. ESLA enhances the spatial semantic features and weak boundary information of camouflaged organisms while suppressing overlapping underwater background noise and texture interference. Consequently, it resolves the issues of inaccurate boundary localization and missed detection in turbid water. Additionally, we introduce RCSR, a residual-based convolution module designed to compensate for ESLA’s limitations in channel semantic discrimination. By enriching and refining fine-grained channel-wise feature representations, RCSR preserves the spatial details enhanced by ESLA. This approach effectively reduces the misclassification of morphologically similar species and cross-scale targets at various growth stages. Experiments are conducted using the UTDAC2020 and Brackish datasets. Experimental results show that EL-DETR achieves 84.4% mAP on the UTDAC2020 dataset and 94.4% mAP on the Brackish dataset, outperforming state-of-the-art underwater object detection models in both detection precision and localization accuracy. The proposed model provides a high-precision solution for automated organism counting, growth status monitoring, and biomass estimation in intensive aquaculture.