Abstract
Infrared imaging offers significant advantages over visible-light imaging in real-world applications, performing robustly under night time, adverse weather, low illumination, and smoky or dusty conditions. However, it suffers from critical limitations, including lack of color information, low signal-to-noise ratios (SNR), and poor object-to-background contrast, which degrade standard object detection algorithms. To address these challenges, we propose YOLO-MDIR, a specialized infrared object detection network built upon the YOLO framework. We introduce a novel multi-directional convolution(MDC) module that adopts a parallel architecture, integrating horizontal, vertical, differential, and standard convolution kernels. This design enables the network to capture structural information along different orientations and extract multi-scale features, significantly improving its perception of infrared objects characterized by weak textures and low contrast. Furthermore, real-world thermal camera operations such as brightness adjustment, contrast stretching, and non-uniform noise can compromise detection stability. To mitigate these issues, we employ data augmentation strategies including random brightness and contrast adjustment, black-white polarity inversion, and non-uniform noise simulation. We evaluate YOLO-MDIR on a hybrid dataset combining the public FLIR and IRay datasets. Experimental results demonstrate that compared with mainstream networks including YOLOv5, YOLOv8, and YOLO11, our proposed YOLO-MDIR achieves improvements in mean Average Precision (mAP) of approximately 5.1%, 3.6%, and 3.2%, respectively. These results validate the superior detection performance of our algorithm in infrared scenarios and its practical utility for complex environments.