An Improved YOLOv8-Based Pest Identification Method for Farmland

In response to the problems of large model parameters and high computational requirements in existing farmland pest identification models, an enhanced pest identification approach based on YOLOv8 for farmland has been put forward. This approach aims to lower the computational burden and memory consumption while enhancing the target detection speed. The approach incorporates a lightweight convolutional module, GhostConv, in the Backbone layer to substitute the common convolution in the original YOLOv8 feature extraction network, thereby reducing the network parameters. Furthermore, a lightweight C3Ghost module is introduced to replace the original C2f module in the feature extraction network, further reducing the computational load. Tests on a constructed farmland pest dataset indicate that compared with the YOLOv8 model, the mAP value of the improved model reaches 90.8%, which is basically equivalent to that of the YOLOv8 model; the computational load is reduced from 8.1 GFLoPs to 5.0 GFLoPs, a decrease of 38.3%. Compared to the current mainstream object detection algorithms, the improved model demands fewer computational resources and is faster without sacrificing detection accuracy, giving it an edge in deployment, especially in scenarios with limited resources or requiring real-time processing. This optimization is highly valuable for applications that aim to reduce computational costs and improve response speed while maintaining high precision.