A multi-scale parallel weighted fusion dynamic attention method for citrus leaf disease recognitions

To address the low detection accuracy caused by leaf occlusion, the loss of disease targets, and complex backgrounds in citrus leaf disease detection, this study proposes a leaf disease detection method termed DBG-DETR (a real-time detection transformer with DMGF, BDFF, and GSDT). Firstly, a DMGF-ResNet18 (dynamic multi-scale gating fusion block) is designed as the disease feature extraction module. By leveraging multiscale parallel depthwise separable convolutions, this module adaptively extracts and fuses rich disease-related features. Secondly, a GSDT (gated sparse dynamic transformer) is introduced to focus on deep features. Through a dynamic gating mechanism and Top-K sparse attention, GSDT reduces model parameters while enabling the network to concentrate on disease regions. Finally, a BDFF (bi-directional dense feature fusion module) is proposed to facilitate effective interaction between shallow and deep features, achieving efficient disease feature fusion. Experimental results on a real or chard dataset demonstrate that, compared with the baseline model, DBG-DETR improves P, mAP mmAP, R and F1 by 3.31%, 3.40%, 4.11%, 3.89% and 3.59%, respectively, while reducing the number of parameters by 3.78 MB. These results indicate that the proposed method significantly enhances disease detection performance in complex background environments and provides reliable technical support for intelligent citrus orchard management.