Joint decoupling and restoration of motion blur and thermal noise in high-speed infrared thermal imaging

Infrared thermal imaging for industrial process monitoring, such as in the hot forming of aluminum alloy ring forgings, is often compromised by a hybrid degradation of severe motion blur and thermal noise. Existing methods struggle to handle such hybrid distortions, especially in separating frequency-linked blur and noise. To overcome these challenges, we propose WSA-DA-DeblurUNet, a novel deblurring network, to our knowledge, that employs a multi-stage restoration strategy. First, the dynamic channel modulation shallow convolution module (DCM-SCM) counteracts detail loss from motion blur by adaptively modulating shallow features to recover high-frequency details. As this process can amplify inherent thermal noise, the statistical attention feature aggregation module (SA-FAM) is introduced to perform a noise-robust fusion through statistical attention, yielding a detail-enhanced feature map. Finally, for high-fidelity reconstruction, the FFT dual attention-asymmetric feature fusion (FDA-AFF) module refines this feature map by independently modulating the amplitude and phase spectra, which correspond to image texture and structure, respectively. This decoupled enhancement achieves precise reconstruction. Our proposed network, trained on the FLIR dataset and tested on both public benchmarks and real ring-forging images, demonstrates superior deblurring performance and strong generalization.