Rice pest detection based on improved ATSS model

ObjectiveTo address issues including data scarcity, low accuracy, and poor real-time performance in rice pest detection, this study aims to construct a specialized dataset and develop an efficient detection method.MethodThe Pest5 dataset was built based on insect-attracting light traps. Within the adaptive training sample selection (ATSS) framework, an improved model, PestDet, was proposed. Improvements included: A combined data augmentation strategy and anchor optimization were adopted to enhance sample diversity and target matching capability; GHM-C and DIoU were used as the classification and regression losses, respectively, to improve robustness and localization accuracy; Inflated convolutions were introduced to reconstruct the feature pyramid for enhanced multi-scale feature perception; The detection head architecture was simplified and the coordinate attention (CA) mechanism was embedded to accelerate inference and strengthen key information extraction.ResultPestDet achieved a detection mean average precision (mAP) of 92.0% and frame per second (FPS) of 40.2 on the Pest5 dataset, surpassing the original ATSS by 7.0 percentage points and 7.0 respectively, and outperformed other mainstream models.ConclusionPestDet demonstrates high accuracy with high efficiency, enabling effective identification of rice pests in complex backgrounds, and provides technical support for intelligent pest monitoring and precision control.