YOLOv8 forestry pest recognition based on improved re-parametric convolution

IntroductionThe ecological and economic impacts of forest pests have intensified, particularly in remote areas. Traditional pest detection methods are often inefficient and inaccurate in complex environments, posing significant challenges for effective pest management. Enhancing the efficiency and accuracy of pest detection under resource-limited conditions has thus become a critical issue. This study aims to address these challenges by proposing an improved lightweight forestry pest detection algorithm, RSD-YOLOv8, based on YOLOv8.MethodsTo improve the performance of pest detection, we introduced several modifications to the YOLOv8 architecture. First, we proposed RepLightConv to replace conventional convolution in HGNetV2, forming the Rep-HGNetV2 backbone, which significantly reduces the number of model parameters. Additionally, the neck of the model was enhanced by integrating a slim-neck structure and adding a Dyhead module before the output layer. Further optimization was achieved through model pruning, which contributed to additional lightweighting of the model. These improvements were designed to balance detection accuracy with computational efficiency, particularly for deployment in resource-constrained environments.ResultsThe experimental results demonstrate the effectiveness of the proposed RSD-YOLOv8 model. The model achieved a Map@0.5:0.95(%) of 88.6%, representing a 4.2% improvement over the original YOLOv8 model. Furthermore, the number of parameters was reduced by approximately 36%, the number of operations decreased by 36%, and the model size was reduced by 33%. These improvements indicate that the RSD-YOLOv8 model not only enhances detection accuracy but also significantly reduces computational burden and resource consumption.DiscussionThe lightweight technology and architectural improvements introduced in this study have proven effective in enhancing pest detection accuracy while minimizing resource requirements. The RSD-YOLOv8 model's ability to operate efficiently in remote areas with limited resources makes it highly practical for real-world applications. This advancement holds positive implications for agroforestry ecology and supports the broader goals of intelligent and sustainable development. Future work could explore further optimization techniques and the application of this model to other domains requiring lightweight and accurate detection systems.