Research on Real-Time Forestry Pest Detection Based on Improved YOLOv5

Small object detection has always been a difficult point in the field of object detection. To achieve better detection performance of forestry pests, this paper proposes Mf-YOLOv5s. Based on YOLOv5s, we replace the PANet with M-BiFPN to explore the importance of different input features and add one more prediction head to enhance the detection of tiny pests. Then we insert the BoTR between backbone and neck to capture global contextual information by using self-attention mechanism. Furthermore, we use Copy-Pasting data augmentation strategy to expand the dataset, which can make the sample distribution evenly. We also add a D-CBAM to neck to explore the role of hybrid attention mechanism in small object detection. The experimental results show that the AP(50) of Mf-YOLOv5s on the test set is 95.3%, which is 2.2% higher than YOLOv5s, the detection precision and recall are 2.9% and 3.1% higher, respectively.