Answers are grounded in this paper and cite their sources.
• Efficient Channel Attention (ECA) enhances lesion feature extraction under occlusion and uneven illumination. • GSConv combined with BiFPN enables efficient multi-scale feature fusion, improving detection accuracy for large and small lesions. • Group_Taylor pruning with CWD-based channel knowledge distillation reduces parameters and computational cost while maintaining high precision. • The EBGC-RTDETR model achieves a 51.5% reduction in parameters, 67.2% decrease in FLOPs, and improves mAP@0.5 by 2.3% and F1 score by 1.23% compared with RT-DETR. To address the limitations in lesion characterization, regional attention, multi-scale feature fusion, and lightweight deployment in existing models, we propose EBGC-RTDETR, an improved architecture based on RT-DETR-r18. First, an Efficient Channel Attention (ECA) module is added after the final BasicBlock in the backbone to enhance lesion representation and alleviate insufficient feature extraction. Second, this study designs a Ghost-Shuffle Convolution (GSConv) based Bidirectional Feature Pyramid Network (GS-BiFPN), which integrates the lightweight GSConv modules with a bidirectional feature pyramid to replace the original encoder neck, enabling efficient multi-scale feature fusion. Finally, Group-Level First-Order Taylor (Group-Taylor) pruning and Channel-Wise Distillation (CWD) are applied to remove redundant structures and reduce computational cost while preserving essential feature representations. Experimental results show that EBGC-RTDETR achieves notable improvements over RT-DETR-r18: Precision increases by 0.6%, Recall by 1.7%, mean average precision (mAP0.5) by 2.3%, mAP0.5-0.95 by 1.6%, and F1 Score by 1.2%. Meanwhile, the number of parameters and computational complexity (GFLOPs) are reduced by 51.5% and 67.2%, respectively. The proposed model achieves both enhanced performance and lightweight design, providing a valuable theoretical and practical reference for machine vision applications in eggplant disease detection.
No takes yet. Share an insight, caveat, or question.
Huo et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: