YOLO-MIF: Improved YOLOv8 with Multi-Information fusion for object detection in Gray-Scale images

Grayscale images, characterized by their single-channel grayscale information, offer a simplified representation that reduces the complexity and cost associated with processing and storage. They find widespread application in various detection tasks across domains such as medical imaging, industrial inspection, autonomous driving, remote sensing imaging, and surveillance security. However, grayscale images face challenges such as limited object discrimination, susceptibility to noise, and uneven brightness due to the absence of color information. Moreover, the real-time constraints in these fields present additional hurdles for object detection using grayscale images. This article tackles these challenges by introducing an enhanced object detection network, YOLO-MIF, which integrates diverse multi-information fusion strategies to enhance the YOLOv8 network. Initially, a technique was introduced to generate pseudo-multi-channel grayscale images, enhancing the network’s channel information and reducing potential image noise and defocus blur issues. Subsequently, network structure reparameterization technology was employed to boost the network’s detection performance without increasing the inference time. Additionally, a novel decoupled detection head was introduced to amplify the model’s expressive power when dealing with grayscale images. The efficacy of the proposed algorithm was evaluated on two open-source grayscale image detection datasets (NEU-DET and FLIR-ADAS). Results indicate that, at similar inference speeds, the proposed algorithm outperformed YOLOv8 by 2.1% and Faster R-CNN by 4.8% in terms of mAP(mean Average Precision), achieving a superior balance between detection efficiency and effectiveness. The code for the algorithm will be made available at https://github.com/wandahangFY/YOLO-MIF.