A lightweight SSV2-YOLO based model for detection of sugarcane aphids in unstructured natural environments

• Published in: Computers and electronics in agriculture Vol. 211; p. 107961
• Main Authors: Xu, Weiyue, Xu, Tao, Alex Thomasson, J., Chen, Wei, Karthikeyan, Raghupathy, Tian, Guangzhao, Shi, Yeyin, Ji, Changying, Su, Qiong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2023
• Subjects: Deep learning; Dense target recognition; Lightweight; Pest detection; YOLOv5; Deep learning; YOLOv5; Dense target recognition; Lightweight; Pest detection
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.107961

[Display omitted] •A lightweight model was developed to detect sugarcane aphids in broad leaf crops.•Reconstructing backbone of You Only Look Once-v5-small reduced model size by 58.5%.•Refactoring neck network width reduced parameter numbers and computing time.•The proposed model outperformed other state-of-the-art algorithms. Accurate, rapid, and smart pest recognition and detection are important for crop protection and management. Existing deep learning-based pest detection algorithms often require high computing resources and computational memory. In addition, detection of small and high-density pests (e.g., sugarcane aphids (SCAs)) in unstructured natural environments still remains challenging due to the small target size, natural illumination, and background noises. Here, we developed a lightweight SSV2-YOLO (Stem-ShuffleNet V2-YOLOv5s) model based on YOLOv5s (You Only Look Once, version 5, small) for SCA detection by reconstructing the backbone network with Stem and ShuffleNet V2 and adjusting the neck network width. We further refactored the feature level, data augmentation method, and loss function to improve detection performance for small, high-density, and overlapping targets. The newly developed SSV2-YOLO model was trained and tested using a mobile phone SCA image dataset (n = 860) covering both uniform and unstructured natural environments. Compared with the original YOLOv5s, SSV2-YOLO significantly reduced model complexity, e.g., reduced parameter number (by 95.1%), model size (by 92.5%), and floating point of operations (by 81.6%), and showed increased detection speed (GPU by 31.6% and CPU by 81.8%) and improved accuracy (by 2.5%). SSV2-YOLO also outperformed the best state-of-the-art algorithms for SCA detection under severe adhesion and unstructured farmland conditions. We found 640 × 640 as the best image pixel resolution to improve model detection performance without affecting model size. Our study showed that SSV2-YOLO has clear advantages in terms of accuracy, efficiency, and lightweight, which can be potentially used for the automatic detection of SCAs on mobile devices. The evaluation code and dataset are available at https://github.com/weiyuexu/SCAs-SSV2-YOLO.