Infrared Image Denoising based on Multi-Level Feature Attention to Achieve Wildfire Detection

Infrared image denoising has significant application value in areas such as forest fire prevention. Existing denoising methods based on deep learning struggle to distinguish between complex noise and weak fire point characteristics, which can lead to excessive smoothing or loss of target information, severely affecting subsequent detection accuracy. To address these issues, this paper proposes an infrared image denoising algorithm based on multi-level feature attention. This algorithm first achieves the separation of high and low-frequency features in infrared images through a dynamic scale-aware convolution, enhancing the thermal background robustness of low-frequency components using an adaptive contrast modulation network; secondly, it designs a multi-scale cross-attention network to guide the reconstruction of high-frequency details with low-frequency features, collaboratively modeling the distribution differences of noise and fire points in the spatial frequency domain, reducing false alarms while preserving target edge characteristics. Experiments demonstrate that this algorithm has significant advantages in the coordinated optimization of noise suppression, detail retention, and computational efficiency.