Prior Guided Multi-Scale Dynamic Deblurring Network for Diffraction Image Restoration in Droplet Measurement

High-precision measurement of micrometer-scale flying droplets is demanded in inkjet printing manufacturing. However, the measurement accuracy is limited by droplet image degradation caused by optical diffraction and actual imaging conditions. Most existing image restoration methods focus on defocus blur and motion blur and pay less attention to diffraction degradation, which cannot handle real-world complex degradation well. In this study, to address the challenges in droplet image restoration, we propose a diffraction-Gaussian degradation framework to simulate actual degradation and a prior guided multiscale dynamic deblurring network (PDDN) for image restoration. PDDN explicitly utilizes degradation prior information with the proposed fast Fourier transform (FFT)-based prior extraction (FPE) module and the multiscale dynamic deblurring (MSDD) module. FPE extracts the degradation prior with the combination of Weiner deconvolution and deep learning. MSDD restores intermediate features using kernel prediction-based dynamic convolution under the guidance of the learned prior. PDDN employs a U-shaped Transformer architecture along with prior guided dynamic deblurring to achieve nonblind deblurring. Experiments on four synthesized datasets demonstrate that PDDN achieves state-of-the-art performance in diffraction image restoration. The effectiveness of the degradation framework and PDDN is proved in real-world image restoration, with droplet measurement accuracy improved from 3% to 2.42%.