Robust rice-crab detection via receptive-field attention and feature reassembly in unstructured agricultural scenes

Accurate and real-time detection of rice crabs in paddy-field aquaculture is crucial for precise feeding and behavior monitoring. This task is challenging due to the high textural similarity between crabs and their environment, partial occlusion from rice blossoms, and difficult environmental conditions such as variable illumination and adverse weather, which lead to feature confusion and detail loss. To address these challenges, we propose YOLO-RCCS, a novel rice-crab detection model based on an improved YOLO11n architecture. We introduce a receptive-field attention (RFA) mechanism to dynamically generate content-aware convolutional kernels, enhancing the model's ability to distinguish target features from background noise. We then augment the model with a coordinate-aware attention (CAA) module to capture long-range spatial dependencies, improving target localization. Furthermore, we replace the original upsampling layer with the CARAFE module, which uses content-aware feature reassembly to mitigate detail loss and preserve local textures. Finally, we employ the SlideLoss function to address class imbalance by dynamically re-weighting hard samples, thereby stabilizing the training process. Experimental results demonstrate that YOLO-RCCS achieves a 2.1% increase in precision, a 4.3% increase in recall, and a 3.5% increase in mAP@0.5 compared to the baseline YOLO11n, while maintaining comparable computational complexity. The proposed model shows promise as a potential solution for all-weather monitoring and precise feeding in rice-crab aquaculture, indicating strong feasibility for real-time applications under laboratory conditions.