FRAISE-Net: Scale-aware attention and adaptive cross-scale fusion for strawberry disease inspection

Early and reliable detection of strawberry foliar diseases is essential for reducing yield loss, limiting unnecessary pesticide use, and supporting timely site-specific crop management. This paper presents FRAISE-Net , a one-stage Strawberry Disease Inspection Network for automated localization and recognition of strawberry disease symptoms. The proposed architecture consists of an enhanced disease-aware backbone, a PAN-FPN-style neck, and a three-scale detection head. It integrates three complementary modules: Scale-Aware Channel Attention (SACA) to emphasize lesion-relevant channel responses, Context Aggregation Module (CAM) to capture broader disease context, and Adaptive Cross-Scale Fusion (ACSF) to combine fine spatial details with high-level semantic features across detection scales. Experiments were conducted on a 7-class strawberry disease detection dataset constructed from KUStraw and SDDD after removing the healthy class. The final experimental split contained 4766 annotated images, including 3810 images for training, 474 images for validation, and 482 images for testing. The FRAISE-Net algorithm was implemented on a real robotic arm for real-time disease detection in a controlled greenhouse laboratory setup. Under a unified 100-epoch training protocol with an input size of 320 × 320, FRAISE-Net was compared with YOLO-v9, YOLO-v10, YOLO-v11, YOLO-v12, YOLO-v26, and RT-DETR. On the held-out test set, FRAISE-Net achieved 95.61% precision, 87.88% recall, 91.58% F1-score, 94.48% mAP 50 , 77.58% mAP 50 : 95 , and 35.98 FPS on an RTX 2070 GPU. Five-fold cross-validation showed low fold-to-fold variability across partitions of the current combined dataset, with mean F1-score of 92.15 ± 0.49 % , mean mAP 50 of 95.91 ± 0.32 % , and mean mAP 50 : 95 of 80.09 ± 0.73 % . Ablation analysis confirmed the complementary contributions of SACA, CAM, and ACSF, with the complete FRAISE-Net achieving the strongest overall performance. The robotic experiments support the feasibility of the approach under the controlled, greenhouse-like laboratory conditions evaluated here; however, its generalizability to different sites, camera systems, seasons, disease distributions, and open-field settings remains to be established. These results indicate strong performance on the evaluated dataset, strong false-positive control, and practical inference speed. The code will be available at GitHub: FRAISE-Net .