Pest detection is vital for maintaining crop health in modern agriculture. However, traditional object detection models are often computationally intensive and complex, rendering them unsuitable for real-time applications in edge computing. To overcome this limitation, we proposed DGS-YOLOv7-Tiny, a lightweight pest detection model based on YOLOv7-Tiny that was specifically optimized for edge computing environments. The model incorporated a Global Attention Module to enhance global context aggregation, thereby improving small object detection and increasing precision. A novel fusion convolution, DGSConv, replaced the standard convolutions and effectively reduced the number of parameters while retaining detailed feature information. Furthermore, Leaky ReLU was replaced with SiLU, and CIOU was substituted with SIOU to improve the gradient flow, stability, and convergence speed in complex environments. The experimental results demonstrate that DGS-YOLOv7-Tiny performs excellently on the tomato leaf pest and disease dataset, with 4.43 million parameters, 10.2 GFLOPs computational complexity, and an inference speed of 168 FPS, achieving 95.53% precision, 92.88% recall, and 96.42% mAP@0.5. The model delivered faster inference and reduced computational requirements while maintaining competitive performance, offering an efficient and effective solution for pest detection in smart agriculture with substantial theoretical and practical value.

Artificial Intelligence(AI); Depthwise separable convolution (DSC); Lightweight model; Mobile edge computing (MEC); Object detection; Depthwise separable convolution (DSC); Lightweight model; Mobile edge computing (MEC); Object detection