MSMD-YOLO: Multi-scale and multi-directional Mamba scanning infrared image object detection based on YOLO

Abstract

Infrared (IR) imaging technology, which operates independently of light and weather conditions and can penetrate clouds and soot, offers unique advantages for object detection. However, detecting objects of varying scales remains a significant challenge due to object size, distance, resolution, and scene complexity differences. To address these challenges, we propose a multi-directional and multi-scale localized feature-enhanced infrared object detection method based on YOLOv7. The proposed model introduces the Mamba module with a selective mechanism and multi-scale feature branching to effectively capture object details at different scales. The S-ELAN module integrates multi-directional scanning with a deep convolutional structure to enhance multi-scale feature extraction. Moreover, the local feature enhancement module expands the receptive field using dilated convolution, improving feature representation through the CBAM attention mechanism. It enhances the model’s semantic understanding of objects. Experimental results on a self-constructed multi-scale infrared object dataset demonstrate that the proposed model adeptly tackles the complexities inherent in detecting objects across various scales. Specifically, experiments on the MSIR dataset revealed an mAP0.5 score of 96.8%, which is 4.4% higher than the baseline model, YOLOv7. Furthermore, on the FLIR public dataset, the proposed model achieves an mAP0.5 score of 86.6%, outperforming YOLOv7 by 4.0%. These findings indicate significant performance improvements over prevalent object detection algorithms, highlighting the model’s effectiveness and strong generalization ability in infrared object detection. The code is available at https://github.com/ELF233/MS-MD-YOLO.