An RT-DETR-Based Strawberry Disease Detection Method via Multi-Domain Collaborative Optimization

Accurate strawberry disease detection under complex field conditions is challenging because lesion appearance and scale vary considerably and disease regions are often affected by occlusion, illumination changes, and complex backgrounds. To improve feature representation and multi-scale information interaction, this study proposes RT-DETR-AGWZ based on RT-DETR-R50. The proposed model introduces BottleNeck_Attention for spatial feature enhancement, GLSA for global–local contextual modeling, WaveletUnPool for Haar-basis-constrained structured upsampling, and Zoom_cat for multi-scale feature alignment and fusion. Experiments were conducted on a cleaned, self-collected strawberry disease dataset using a frozen group-aware data partition. Across three independent runs on the validation set, RT-DETR-AGWZ achieved a mean Precision of 0.948±0.006, Recall of 0.944±0.006, [email protected] of 0.969±0.002, and [email protected]:0.95 of 0.836±0.008. Compared with RT-DETR-R50, the mean [email protected]:0.95 increased by approximately 1.93 percentage points, from 0.817±0.012 to 0.836±0.008. Error analysis further showed reductions in false-positive detections and localization errors, and RT-DETR-AGWZ achieved the best overall detection performance among the representative detectors evaluated under the same experimental protocol. These results indicate that the proposed multi-domain collaborative optimization strategy improves strawberry disease detection and localization under complex field conditions.