Answers are grounded in this paper and cite their sources.
To address the challenges of small lesion scales, overlapping leaf occlusion, and complex background interference in citrus leaf disease detection, this study proposes MSAF-DETR, a detection method specifically designed for citrus leaf diseases. First, an MSGC-C2f-ResNet18 backbone is developed for feature extraction. By embedding a multi-scale directional feature scanning module into the C2f stages of C2f-ResNet18, horizontal, vertical, and diagonal scanning mechanisms are employed to strengthen the representation of small lesions and disease regions affected by occlusion. Subsequently, a lightweight global context modeling module, termed SDTA, is introduced. It combines depthwise convolution with cross-channel transposed self-attention to enhance disease-related responses while suppressing irrelevant background interference. In addition, an Adaptive Cross-Scale Modulation (ACSM) fusion module is constructed to achieve effective alignment and integration of shallow spatial details with deep semantic features through dynamic local relation refinement and cascaded dilated convolution. Finally, the Focal-DIoU loss is introduced to reduce localization errors during bounding-box regression. Experimental results on the public CLS dataset and the self-constructed JXCL citrus leaf disease dataset show that MSAF-DETR achieves mAP values of 88.74% and 99.23%, respectively, representing improvements of 6.94% and 3.52% over the baseline RT-DETR model. The proposed method also maintains high detection accuracy and robustness under challenging orchard conditions, including uneven illumination, branch and leaf occlusion, and small-scale disease lesions, providing an efficient and reliable solution for the visual perception systems of citrus-picking robots.
No takes yet. Share an insight, caveat, or question.
Lu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: