PestRefineDet: integrating DenseNet with RefineDet for precision pest detection in agriculture.

Introduction: Insect pests are a major threat to agricultural productivity, causing significant crop losses and reducing food production. Accurate and timely pest detection is essential for minimizing economic losses and supporting sustainable crop management. However, manual pest identification is labor-intensive, time-consuming, and challenging due to the high visual similarity among many pest species and the limited availability of expert knowledge. Methods: To address these challenges, we propose PestRefineDet, a deep learning-based framework that integrates DenseNet-41 with RefineDet for automated pest detection and classification. DenseNet-41 is employed as the backbone network to enhance feature extraction through dense connectivity, enabling effective learning of fine-grained pest characteristics. The extracted features are subsequently processed by the one-stage RefineDet framework for simultaneous pest localization and classification. Results: The proposed framework was evaluated on the challenging IP102 benchmark dataset. Experimental results demonstrate that PestRefineDet achieves a mean Average Precision (mAP@0.5) of 83.61%, providing accurate localization and reliable classification of diverse pest species while maintaining computational efficiency. Discussion: The results indicate that the proposed framework effectively captures discriminative pest features and improves detection performance on a large-scale, real-world pest dataset. PestRefineDet provides a promising solution for automated agricultural pest monitoring and has the potential to support intelligent and sustainable pest management applications.