YOLO-Mamba: object detection method for infrared aerial images

At present, unmanned aerial vehicle (UAV) is widely used in various application fields. Through the detection of nighttime infrared images taken by UAV, it is convenient to analyze the ground situation in real time. Due to the problems of image blur and image noise in infrared images, higher requirements are put forward for object detection algorithms. Aiming at the problems of long-distance dependence and computational complexity of current object detection algorithms based on CNNs and self-attention mechanism, a new infrared aerial object detection method named YOLO-Mamba was proposed. The method combines Mamba with the attention mechanism, a new attention module based on Mamba was proposed. It uses Mamba to scan the features of the feature dimension and the spatial dimension of the image, and fully extracts the global context information, which further improves the algorithm's attention to the key area of the image and reduces the influence of redundant information. In the experiment, through the public infrared aerial image data set, the effectiveness of the improved attention module is verified from both quantitative and qualitative perspectives. The experimental results show that compared with SE and CBAM attention mechanism, the mAP50 is increased by 0.8% and 1.3% respectively, and the parameter quantity is between the two. The attention to critical regions is much higher than other algorithms. Finally, compared with other object detection method, the mAP50 is increased by 1.1% and the Map50-95 index is increased by 0.8% compared with the benchmark model YOLOv8n, and the number of parameters is only increased by 0.1 M. The research results provide a certain reference for improving the accuracy of object detection.