Infrared Small Object Detection Based on Spatial Distribution Fusion and Multi-Scale Upsampling

Detecting small infrared objects is a challenging task due to the inherent limitations of infrared images, such as insufficient texture information and low spatial resolution. To enhance detection accuracy, we propose an Original Image Attention Module(OIAM), which captures the spatial distribution patterns of objects within the dataset, and integrates the learned spatial distributions with the original image through a spatial attention mechanism. The output of OIAM is fed to the backbone of YOLOv8 to obtain efficient features for object detection. Additionally, we introduce a Multi-Scale Upsampling Module(MSUM) that fuses low- and high-level features during the upsampling phase, further enhancing the features of small infrared objects without increasing computational complexity. Experimental results on the FLIR dataset demonstrate that our method effectively improves the detection accuracy of small infrared objects, achieving a 3% increase in overall mAP@0.5 and a 7.2% improvement in mAP@0.5 for bicycles, which primarily consist of small objects.