Research on Small Object Detection in Vehicle-Mounted Infrared Images Based on YOLO-OS
-
Abstract
Owing to the inherently low contrast of vehicle‑mounted infrared imagery—which renders target details and contours indistinct, leading to frequent loss of small‑target features and elevated misclassification rates—we present YOLO‑OS (Object‑Small), a novel infrared‑small‑target detection framework. First, we introduce an ultra‑small‑target detection head that seamlessly fuses deep semantic features, thereby bolstering the network’s capacity to represent and parse fine-grained information; concurrently, we embed an ADown downsampling convolution to strike an optimal trade‑off between model compactness and accuracy. Next, our backbone incorporates a C3SCC module—integrating spatial‑channel reconstruction convolutions—that elevates detection performance on diminutive targets amidst complex backgrounds while preserving a minimal parameter footprint. Finally, we replace conventional upsampling with a Dynamic Upsampling (Dysample) scheme to enhance interpolation precision during multi‑scale feature aggregation. Experimental evaluation demonstrates that, relative to the YOLOv11n baseline, YOLO‑OS achieves a 38 % reduction in parameter count and boosts mAP from 51.5 % to 57.5 %, evidencing its superior feature‑extraction prowess and detection accuracy for infrared small‑target scenarios.
-
-