PestDet: A unified detection framework for accurate and efficient stored-grain pest detection

Abstract

Integrated pest management (IPM) is essential in the agriculture industry to ensure food safety and quality. Detecting stored-grain pests on the surfaces of grain piles is important in IPM to minimize postharvest storage losses. Recently, numerous deep learning-based detection methods have been proposed. However, accurate perception of morphological features of small-size pests still suffers from various challenges. To address these issues, we propose PestDet, a unified detection framework for accurate and efficient detection of stored-grain pests. Specifically, we propose an enhanced feature extraction block (EFEB) with a large effective receptive field (ERF) and integrate it into a designed backbone network, PestBak. Thus, rather than solely focusing on texture features of a target, the model can also focus on more detailed features regarding the shape and contour of small pests, compared to networks with smaller ERFs. Meanwhile, we also present a one-to-many label assignment (OMLA) strategy for accurate feature perception to effectively mitigate the imbalance between the number of positive and negative samples by assigning more positive samples in the training phase. In addition, it adeptly handles the uncertain assignments of the samples. Furthermore, a regression loss based on normalized Gaussian Wasserstein distance (NWD) is designed to improve detection accuracy and model convergence by introducing an additional penalty for the location deviation of the predicted bounding boxes. In addition, Reparameterization is integrated to accelerate the inference speed. Extensive experiments are conducted on GrainPest, a scenario-based dataset. PestDet achieves state-of-the-art performance with a mAP of 90.6 %, precision of 85.6 %, and recall of 88.0 %, indicating that it can serve as a general pipeline for pest detection aimed at monitoring stored-grain pests in granaries. Our code and data are available at (https://github.com/IntelligentsystemlabTian/PestDet).