MBFormer-YOLO: Multibranch Adaptive Spatial Feature Detection Network for Small Infrared Object Detection
Xiao Luo, Shaojuan Luo, Meiyun Chen, Genping Zhao, Chunhua He, Heng Wu
Abstract
Infrared thermal imaging technology that has unique advantages in obtaining target information at night and in harsh weather conditions has been widely used in the military, security, environmental monitoring, and other fields. Existing visible light object detection methods often face challenges of high false alarm rates and low accuracy when detecting small infrared objects. Moreover, current small infrared object detection methods mostly require much detailed annotation data, and it is challenging to visually depict the positions of small targets on the original complex background images. To address these issues, we propose a multi-branch backbone and adaptive spatial feature fusion detection head-based YOLO network called MBFormer-YOLO to achieve the high-accuracy detection of small infrared targets. We design a multi-branch backbone with structural re-parameterization capability called MBFormer, which has strong capabilities to fit object features. In the neck part, MBFormer-YOLO utilizes convolutional block attention modules and C3-SwinTransformer modules to suppress noise and redundant information. We develop a 4-level adaptive spatial feature fusion detection head to improve the detection accuracy of the proposed network. Many experiments are conducted on the SIRST-V2 dataset to validate the effectiveness of MBFormer-YOLO. The results show that MBFormer-YOLO achieves an 11.3% increase in AP50 and 2% in AP50-95 over the baseline model YOLOv8n and surpasses other advanced object detection models.