Wavefront coding image reconstruction based on learnable physical prior and frequency domain attention block

Wavefront coding (WFC) is an effective technique for extending depth-of-field of imaging systems, including optical encoding and digital decoding. This paper uses the genetic algorithm to design a phase mask in a large field-of-view fluorescence microscope system. And we applied physical prior information and frequency domain model to the WFC decoding, proposing a reconstruction method by a generative model. Specifically, we rebuild the down-sampling and up-sampling modules inspired by the Transformer to improve the model's overall performance. We propose the point spread function (PSF) attention layer, which explicitly learns the deviation between the out-of-focus PSF and the on-focus PSF. Meanwhile, the multi-feature fusion block is used to ensure which PSF feature information should be retained and reasonably integrated into the details and semantic information of the image. In addition, we constructed the frequency domain self-attention block to separate further and reconstruct the PSF feature information, the high and low frequencies of the image in the frequency domain. The experimental results show that the decoding model effectively reduces noise, artifacts, and blurring compared to other methods. Therefore, we construct a deep-learning WFC optical computational imaging method at visible wavelengths, with good potential in detecting Digital PCR (dPCR) and biological images.