Abstract
In tomato cultivation, various diseases significantly impact tomato quality and yield. The substantial scale differences among diseased leaf targets pose precise detection and identification challenges. Moreover, early detection of disease infection in small leaves during the initial growth stages is crucial for implementing timely intervention and prevention strategies. To address these challenges, we propose a novel tomato disease detection method called TomatoLeafDet, which integrates multi-scale feature processing techniques and small object detection technologies.Initially, we designed a Cross Stage Partial -Serial Multi-kernel Feature Aggregation (CSP-SMKFA) module to extract feature information from targets at different scales, enhancing the model's perception of multi-scale objects. Next, we introduced a Symmetrical Re-calibration Aggregation (SRCA) module, incorporating a bidirectional fusion mechanism between highresolution and low-resolution features. This approach facilitates more comprehensive information transmission between features, further improving the efficacy of multi-scale feature fusion. Finally, we proposed a Re-Calibration Feature Pyramid Network with a small object detection head to consolidate the multi-scale features extracted by the backbone network. This network provides richer multi-scale feature information input for detection heads at various scales. Results indicate that our method outperforms YOLOv9 and YOLOv10 on two datasets. Notably, on the CCMT tomato dataset, the proposed model achieved improvements in mean Average Precision (mAP50) of 4.4%, 1.9%, and 2.3% compared to the baseline model, YOLOv9s, and YOLOv10n, respectively, exhibiting significant efficacy.