An Infrared Dim-small Target Detection Method Based on Improved YOLOv7

Abstract

Efficient detection of dim-small targets with high accuracy is a difficult task in the field of infrared target tracking since the tiny size of small infrared targets significantly reduces the accuracy of conventional models. To address this issue, this paper improves YOLOv7 so that it can be applied to the detection of infrared dim-small targets. Initially, an enhanced MPConv-based pooling structure is proposed, which reduces the high false detection rate caused by white point noise. Then, a CBAM attention module is added to the backbone structure, which employs both spatial and channel attention to preserve more of the original characteristics of infrared faint targets. Finally, the EIOU loss is utilized in the Head module to increase the speed of model convergence. Experiments reveal that the improved algorithm achieves a model mAP of 70.8% on the dim-small target dataset, which represents a 3.4% improvement over YOLOv7 and outperforms other conventional algorithms.