Improved YOLOv5 for Small Pest Detection: Algorithm Design and Performance Evaluation

Detecting pests in the field is challenging, mainly because most of the targets are quite small — more than half of them — and they vary a lot in size. While deep learning models like DCNN often work well in lab environments, they don’t adapt as easily when applied to real-world farm images, because they struggle with objects of very different sizes. In this work, we try to improve how well YOLOv5s can detect small pests in crop images. We made three main changes: replacing the Neck with a CSConv-based module, adding the SimAM attention module, and using Normalized Wasserstein Distance (NWD) loss to better handle small-object prediction and keep things fast. The dataset used in this study includes images from 24 common agricultural pest categories and uses the VOC annotation format and widely used to test object detection models in precision agriculture. Tests on pest datasets showed that our model worked noticeably better, with the [email protected] score going up by 12.1%. It can recognize 12 types of pests commonly found in crops. For example, it does well on Ostrinia furnacalis and Gryllotalpa orientalis. For small pests like Agriotes fuscicollis, the improvement was even greater — about 20% higher than before. Because of the model size at the MB level and inference speed at the ms level, it could be used in many real-world crop monitoring tasks. The model also has limits. Under difficult conditions like occlusion or low image quality, false detections and missed detections can happen. We hope this method can be a useful tool in the shift toward digital agriculture and help with smarter pest management in farming.