YOLO-LDIST: Improved YOLOv9 Long Distance Infrared Small Target Detection

The detection task of infrared small targets from the perspective of long-distance drones is challenging due to weak features, low contrast, and easily interfered by noise. In this paper, based on YOLOv9, we propose an improved infrared small target detection model YOLO-LDIST. As the use of depth wise convolution in YOLOv9 will cause information loss because of the low resolution of infrared small targets, a Space-to-Depth-Batch Normalization and SiLU module (SD-BNS) is proposed in order to reduce the loss of shallow information, reduce the model complexity and computation and enhance the nonlinear expression of features. In addition, we design an Adaptive Channel Attention mechanism (ACA-Net) that allocates more attention to small targets with few pixels in infrared images. The ACA-Net is integrated with Generalized Space Convolution (GSConv) to lightweight neck network, improving computational efficiency. Experiments on infrared aerial datasets show that the proposed method achieves 92.4% and 63.2% of mAP@0.5 and mAP@0.5:0.95, which are 1.2% and 2% higher than the original YOLOv9, respectively. At the same time, the proposed method has lower GFLOPS and can be operated lightly.