HCC-RT-DETR: Residual Channel Calibration for Small-Lesion Detection in Tobacco Field Images
Year: 2026
Type: article
Abstract: Tobacco leaf lesions vary in size, appearance, and spatial distribution, making detection difficult under field conditions. We developed HCC-RT-DETR (Hierarchical Channel Calibration for RT-DETR), an RT-DETRv2-based detector for four infectious disease categories, the abiotic disorder Weather Fleck, and a Healthy reference class. The module derives channel weights from globally averaged features using a local one-dimensional convolution and sigmoid activation. Learnable residual terms add the reweighted features to the original representations. Calibration strength α is initialized at 0.01 for the high-resolution P3 feature and 0.05 for the P4 and P5 features. On the held-out test split of 780 field images, HCC-RT-DETR achieved a three-seed mean of 64.35% [email protected]:0.95 and 28.50% AP_small. These values exceed the baseline by 1.46 percentage points (2.3% relative) and 10.32 percentage points (56.8% relative), respectively. HCC incurred a modest latency increase on the CPU: 1.64% under ORT FP32 and 0.97% under ORT FP16-converted execution. A validation-based sensitivity analysis examined four initializations of P3 calibration. Evaluation was limited to Yunyan 301 at the vigorous growth stage, using clear-weather images collected between 10:00 and 14:00; transfer to other cultivars, seasons, or weather conditions remains untested.
Source: Agriculture
Institution: Yunnan Agricultural University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7225316474