MDP-YOLO: A LIGHTWEIGHT YOLOV5S ALGORITHM FOR MULTI-SCALE PEST DETECTION

ABSTRACT

Rice pest detection technology plays a crucial role in enabling food production, ensuring ecological balance, supporting sustainable agriculture, and promoting the health of farmers. However, existing technology is faced with challenges such as low detection accuracy, high computational complexity, and large model sizes, making it unsuitable for mobile deployment. This paper presents MDP-YOLO, a lightweight rice pest detection model based on the YOLOv5s model. It includes improvements such as the use of ShuffleNetV2 as the backbone network to significantly reduce the number of parameters and complexity of the model; the introduction of GhostConv to replace redundant convolutional layers, in order to further reduce the computational complexity and size; the integration of a large-scale feature extraction layer to enhance the ability of the algorithm to detect small rice pest objects; and the use of CBAM to increase the focus on the regions of interest. Tests on collected datasets show that the modified MDP-YOLO model increases the mean average precision by 5.5% and reduces the FLOPs, parameters, and model size by 72.4%, 89.2%, and 70.1%, respectively, compared to the original YOLOv5s model. Ablation experiments indicate that MDP-YOLO gives superior detection rice pest detection performance compared to other algorithms. MDP-YOLO can effectively detect rice pests, and provides theoretical and technical support for the deployment of lightweight rice pest detection models in practical scenarios.

deep learning; YOLOv5s; object detection; pest detection; lightweight; attention mechanism