S-RPN: Sampling-balanced region proposal network for small crop pest detection

Abstract

Effective pest management and control are the key factors in the agricultural food safety field. Therefore, the automatic monitoring and precise recognition of crop pests have a high practical value in the process of agricultural planting. Over the years, pest recognition and detection results have been rapidly improved with the development of deep learning-based methods. Although promising, these methods still have limited efficiency and precision to detect crop pests with very small scales, deteriorating their effectiveness. The main reason is that current deep learning-based methods may not be able to extract sufficient detailed appearance features for small pest objects in an image, making it difficult to train a classifier to detect and distinguish small pests from the backgrounds or similar objects. To address the small pest recognition and detection problem, in this paper, we instead seek to recast the current region proposal network and perform more details in different scales for easier small pest detection. Inspired by the visual attention system, we first introduce attention mechanism into the Residual network for obtaining richer pest feature appearance, especially the detailed features of small object pests; Then, to make the region proposal network (RPN) obtain more high-quality object proposals for easier detection, a sampling-balanced region proposal generation network is proposed for improving pest detection accuracy. Furthermore, we devise a novel adaptive region of interest (RoI) selection method to learn features from different levels of the feature pyramid. Several experiments were conducted on the proposed AgriPest21 dataset, and our method can achieve an average recall of 89.0% and mAP of 78.7%, outperforming other state-of-the-art methods, including SSD, RetinaNet, Free-Anchor, PISA, Grid RCNN, and Cascade RCNN detectors.