IYOLO: Multi-scale infrared target detection method based on bidirectional feature fusion

Abstract

Convolutional neural networks are advanced computer vision solution for various tasks, and have made considerable progress. In order to detect infrared targets with details and texture features, we propose a detection method based on YOLOv3. First, we propose a bidirectional feature fusion structure called Improve-FPN(Im-FPN), which allows fast and efficient multi-scale feature fusion. Second, we use the Focal Loss(FL) loss function to balance the problem of sample imbalance. Based on these optimizations, an Infrared-YOLO (IYOLO) target detection model for infrared target is constructed. On the self-constructed infrared dataset: Infrared-VOC, the mAP of IYOLO can achieve 77.1%, which is 4.0% higher than YOLOv3, and the detection speed can reach 55.6FPS. Experimental results show that this method can improve the overall accuracy of detection while ensuring real-time detection.