| Summary: | Effective and timely detection of citrus diseases and pests is crucial for preserving crop health, optimizing agricultural yield, and mitigating economic losses. Although deep learning has greatly improved the accuracy of pest and disease identification, small pest and disease targets, and complex citrus orchard scenarios have brought great challenges to pest and disease identification. To solve these difficulties, the Light-SA YOLOV8 (Lightweight Self-Attention YOLOV8) model is introduced, tailored specifically for real-time detection within intricate natural settings. The model addresses challenges presented by complex backgrounds, such as uneven lighting and reflections on citrus leaves and fruits, by incorporating the BRA self-attention mechanism module before the SPPF layer in the backbone, while achieve flexible computation allocation and content awareness. Furthermore, to streamline computational complexity, the backbone incorporates the FasterNet Block. Moreover, to elevate precision and computational efficiency in the detection of citrus diseases and pests, an innovative feature fusion technique known as the AFPN (asymptotic characteristic pyramid network) structure is introduced at the Neck. The constructed dataset includes five types of diseases: anthracnose, citrus canker, melanosis, scab, and bacterial brown spot-along with one type of insect pest, the citrus shallow leaf moth. The experimental findings showcase the Light-SA YOLOV8 model’s robust performance, boasting an average detection accuracy of 92.6% across the six disease and pest categories within the test dataset. Achieving an mAP@0.5 of 92.5%, the model swiftly detects individual images in a mere 3.4ms, while consuming only 4.5MB of memory. Notably, in comparison to the original YOLOV8n, the Light-SA YOLOV8 model demonstrates remarkable enhancements in both detection accuracy and computational efficiency. It showcases a 2.8% surge in precision (P), a 0.9% improvement in mAP@0.5, and a substantial 20.7% reduction in computational workload. Additionally, when contrasted with Faster RCNN, YOLOV3-tiny, YOLOV8n, and YOLOV5n, the Light-SA YOLOV8 model exhibits an average increase in detection precision of 8.8%, 6%, 2.8%, and 1.8%, respectively. These results underscore the model’s superior performance in precision and computational efficiency compared to existing models. This research offers valuable insights for real-time plant pest and disease detection in natural environments with unstructured backgrounds.
|
|---|