Abstract

Reliable vision-based crop pest detection is crucial for precision agriculture to prevent economic losses. However, classifying and localizing tiny insect instances is incredibly challenging. In this work, we present an enhanced YOLOv8 model specially tailored for insect pest detection in agriculture. We propose enhancing the popular YOLOv8 object detector with convolutional block attention modules (CBAM) and squeeze-and-excitation (SE) blocks to boost feature representations of small insect pests amidst complex field images. CBAM and SE blocks incorporate attention mechanisms to focus convolutional filters on the most informative regions. We integrate these modules within YOLOv8’s feature extractor backbone for heightened awareness of sparse, tiny pest targets. We conduct experiments on a real-world agriculture insect pest dataset containing images collected under various practical field conditions. Custom-made pest dataset was used to train YOLOv8 models incorporated with CBAM and SE Net. Quantitative results showed state-of-the-art pest detection accuracy, with a mean average precision (mAP) of 94.5% and an F1 score reaching 0.91 for the model utilizing both CBAM and SE Net. By integrating these additional modules, YOLOv8 leveraged feature recalibration and scaling to achieve improved performance on this pest detection task. The attention modeling provides noticeable gains in accurately localizing clustered and occluded insects while reducing false detections. Our improved Attention augmented-YOLOv8 model strikes an optimal balance between accuracy gains and on-device computational constraints for viable insect pest monitoring systems to enable sustainable agriculture.