To improve the object detection performance of intelligent sorting scenarios for nuclear waste, an object detection network YOLOv10‐RSS (receptive field adaptive convolution RFAConv, SimAM, Shape‐IoU) based on dynamic receptive field and shape‐aware Intersection over Union (IoU) is proposed using YOLOv10. In terms of architecture design, traditional convolutional layer stacking is abandoned, and RFAConv structure is introduced to optimize feature extraction, significantly improving feature representation and detection accuracy. Simultaneously integrating SimAM attention mechanism enables the network to adaptively focus on key areas, suppress background interference, and achieve precise recognition and localization. In terms of training optimization, the Shape‐IoU loss function is used to integrate area overlap and shape similarity information, improving training efficiency and model convergence speed. Verified by the COCO dataset and a custom nuclear waste dataset, the mean average precision (mAP) of YOLOv10‐RSS increased by 3.5% and 4.3%, respectively, compared to the original model. In summary, the improved model has significantly improved detection accuracy, surpassing mainstream YOLO series models in core indicators such as parameter quantity, computational complexity, and accuracy. It has successfully achieved a balance between lightweight and high‐performance, providing a reliable technical solution for resource‐limited nuclear waste intelligent sorting scenarios.
Zhang et al. (Thu,) studied this question.