Lightweight weed detection using re-parameterized partial convolution and collection-distribution feature fusion

Weed detection plays a pivotal role in precision agriculture, enabling targeted spraying to minimize environmental impact. However, the similarity between crops and weeds, coupled with varying growth stages, poses significant challenges. Existing models often suffer from high computational costs and information loss during feature fusion. To address these issues, we propose a lightweight weed detection model named DETR-RPC-CDF. This model employs re-parameterized partial convolution (RPC) in the backbone network to reduce computational redundancy and enhance detection speed. Additionally, we introduce a collection-distribution feature fusion (CDF) mechanism to reduce information loss during feature integration and preserve multi-scale details. Extensive experimental results on the Fine24 dataset demonstrate that our model outperforms mainstream methods, achieving a 2% increase in mAP@0.5 while reducing parameters by 40%, FLOPs by 43%, and increasing detection speed by 40%. Our findings underscore the potential of DETR-RPC-CDF for real-time, lightweight, and accurate weed detection in agricultural applications. The source code of the work is available at https://github.com/wenbin-zheng/DETR-RPC-CDF.