Tomato leaf diseases pose significant threats to crop yield and food security. However, in real-world cultivation environments, factors such as fluctuating illumination, varying leaf occlusion, and ambiguous lesion morphology often compromise detection accuracy. This paper presents the Gradient-aware Bidirectional Retentive Detection Transformer (GBR-DETR), a model designed for high-precision, real-time disease detection. This model is composed of two network structures and a retentive feature aggregation module: (1) a Multi-scale Gradient-Aware Transfer Network (MGAT-Net) is designed to encode gradient information through the Sobel operator, thereby enhancing the localization stability for small and blurry lesions; (2) a Bidirectional Context Pyramid Network (BCPN) is proposed to enable bidirectional interactions among multi-level features through a top-down and a bottom-up pathway, thereby generating multi-scale lesion features and bridging cross-scale semantic gaps; and (3) a Retentive Feature Aggregation Module (RFAM) is used to suppress background noise and establish global feature correlations, thereby enhancing the overall representation capability for lesion recognition. Experiments on the Multi-scenario Tomato Leaf Disease (M-TLD) dataset show that GBR-DETR yields gains of 3.12, 4.88, and 3.41 percentage points in mAP50–95, mAP50, and mAP75, respectively, over the baseline RT-DETR, while also outperforming representative DETR-based and CNN-based detectors. The model demonstrates robust generalization on the PlantDoc cross-domain benchmark, achieving a 2.11% improvement in mAP50 over the baseline. Deployed on the NVIDIA Jetson Orin Nano with TensorRT FP16, it achieves 54 ms latency, enabling real-time disease monitoring on edge devices. This solution provides effective technical support for real-time disease monitoring in smart agriculture.
Zhuo et al. (Fri,) studied this question.