Abstract
Due to the large size of the existing corn seedling and weed detection models, they are difficult to deploy on resource-constrained platforms such as miniature weeding robots. Therefore, it is necessary to achieve a balance between detection accuracy and computational efficiency for the detection models. This paper proposed a lightweight detection model, RT-DETR-DSA, based on the RT-DETR-r18 model. Firstly, a dual-path convolution-attention block was designed to construct the backbone network, which improved the model’s computational efficiency by parallelizing the convolution and attention mechanisms. Secondly, based on the cross-scale feature interaction layer of the original model, this paper proposed the adaptive feature fusion block to complement and enhance the high and low-level features to improve the model’s detection accuracy. Finally, to further reduce the model’s computational load, a sliding-window self-attention block was introduced into the neck network, which divided the feature map into multiple non-overlapping windows and calculated the attention within each window separately. Experimental results showed that the proposed RT-DETR-DSA model was constructed with only 13.0M parameters, corresponding to a 35.6% reduction compared with the original RT-DETR-r18 model; its computational load was measured as only 34.9G, achieving a 40.6% reduction over the original model; the mAP@0.5 of the improved model was 1.2 percentage points higher than that of the original model; and its FPS (frames per second) reached 85.4. The RT-DETR-DSA model proposed in this paper improved computational efficiency while maintaining high detection accuracy in corn seedling and weed detection, and provided a potential technical solution for the visual perception system of miniature weeding robots.
Funding Information
  • Key Technology Research and Development Program of Shandong Province (2025TSGCCZZB0802)
  • China Scholarship Council (202401040020)

This publication has 30 references indexed in Scilit: