IRMSD-YOLO: Multiscale Dilated Network With Inverted Residuals for Infrared Small Target Detection

Infrared small target detection (IRSTD) has seen growing applications in practical scenarios. In recent years, deep learning technologies have achieved significant progress in this field. However, existing algorithms fail to fully account for the distinct characteristics of infrared small targets and inadequately utilize features at different depths. Additionally, challenges stem from the weak distinction between the targets and the background, as well as the interference of complex noise. To address these issues, we designed an inverted residual multiscale dilation (IRMSD-YOLO) network. During the feature extraction stage, we designed an inverted residual feature extraction (IRFE) module to more effectively capture the characteristic information of infrared small targets under complex noise conditions. In the feature fusion stage, we conducted a detailed analysis of the importance of features from different layers. By fusing features from various layers at different frequencies, we effectively achieved frequency-weighted feature fusion (FWFF), balancing target localization information and noise suppression. Moreover, to overcome the interference caused by high-brightness noise in the background and the limitations of local contrast enhancement methods, we developed an inverted residual multiscale dilated (IRMSD) attention module. This module employs efficient feature transformations and multiscale dilated attention (MSDA) to reduce the impact of background noise, expand the receptive field, and improve the perception of small targets, thereby addressing the shortcomings of local contrast methods in complex backgrounds. Finally, recognizing the scarcity of infrared small target datasets, we constructed a real-world infrared small target dataset for complex scene (RISTCS). Through extensive experiments on multiple datasets, the results demonstrate that our proposed method exhibits advanced performance.