M-YOLO: an object detector based on global context information for infrared images

Object detection is an important task in computer vision. While visible (VS) images are adequate for detecting objects in most scenarios, infrared (IR) images can extend the capabilities of object detection to night-time or occluded objects. For IR images, we proposes an infrared object detector based on global context information. Combined with the lightweight network (MobileNetV2) to extract features, therefore the detector is named M-YOLO. Then, dedicated to enhancing the global information perception capability of the model, this paper proposes a global contextual information aggregation model. To preserve multi-scale information and enhance expressiveness of features, a top-down and bottom-up parallel feature fusion method is proposed. Only two detection heads are used to implement a lightweight model, which improves detection accuracy and speed. We use the self-built IR dataset (GIR) and the public IR dataset (FLIR) to verify the superiority of the model. Compared with YOLOv4 (78.1%), the average accuracy of M-YOLO (83.4%) is improved by 5.3% on the FLIR dataset. The detection time (4.33 ms) is less, with a detection speed of 30.6 FPS. On the GIR dataset, the detection accuracy (76.1%) is 6.4% higher than that of YOLOv4 (69.7%), and the detection time (6.84 ms) is lower. Our method improves the performance of IR object detection. The method is able to detect IR ground targets in complex environments, and the detection speed meets the real-time requirements.