Fresnel Zone Aperture Lensless Imaging By Unsupervised Deep Learning

With the continuous pursuit of the functions and performance of imaging systems,lensless imaging systems with high performance,low energy consumption,low cost,miniaturization,and light weight have attracted more and more attention from researchers.Although the image quality of the lensless camera cannot match that of the traditional lens camera,its almost flat shape can fill the gap in the miniature camera market.It has potential practical value in novel applications such as augmented reality and human-computer interaction.So far,the biggest problems in the lensless imaging system from the perspective of encoding and decoding: On the one hand,the seven-dimensional plenoptic function of the real light field can only obtain simple two-dimensional intensity information after passing through the coding modulator,so there is a certain physical information bottleneck.On the other hand,how to decode and reconstruct more dimensional and more flexible original light field information in the signal dimension still faces challenges in computational reconstruction methods.Besides,in order to meet the needs of modern new applications,it remains to be studied to design a more simple,suitable and easy-tooperate lensless imaging system.This research is mainly based on the lensless imaging reconstruction of the Fresnel encoding mask,with the goal of improving the reconstructed image quality of the lensless encoding imaging technology.The specific contents as following:(1)Aiming at the problem of poor reconstruction signal-to-noise ratio and low imaging resolution caused by inherent twin images in lensless imaging systems,this paper proposes an unsupervised deep learning algorithm for the lensless encoding imaging method.The method models the gradient of the image data distribution and trains a score-based matching network to learn good image priors to solve the inverse problem of lensless imaging.Then,in the reconstruction process,the "predictorcorrector" sampling method is adopted,and the high-quality reconstruction of low signal-to-noise ratio images is achieved by alternately iteratively updating the numerical solver and the data consistency item.To verify the effectiveness of the method,different scene datasets are verified in this paper.The objective evaluation indicators and visualization of the reconstructed images have obtained satisfactory results.Thus,the proposed algorithm can effectively eliminate twin image artifacts of the imaging system.(2)Aiming at the requirement of thinner and simpler imaging equipment,this paper optimizes the lensless encoding imaging process.Firstly,in the encoding process,the encoding system uses a single Fresnel Zone Aperture as the encoding mask to realize light field modulation under incoherent light illumination,and there is no special calibration between the encoding mask and the sensor.Secondly,in the decoding process,this paper simultaneously inputs the RGB three channels into the trained network for reconstruction,reducing the complexity of the operation.Additionally,experiments on the performance parameters of the imaging system tested to show that the field of view and resolution results are perform well.(3)There are obvious “grid traces” in lensless decoding and reconstruction of images,and the texture details of the restored image are still not ideal.This paper proposes a reconstruction method of joint total variation model,which smooths the image through gradient descent flow to further improve the quality of image reconstruction.Meanwhile,in order to verify the performance of the proposed model,this paper also explores the relationship between image quality and the number of discretization steps at different noise scales.Experiments show that the model framework proposed in this paper has a flexible structure,stable training,excellent performance in modeling data distribution,and can obtain high-quality samples.Therefore,the advantages of this research method in terms of imaging quality and practicability are expected to be applied to other computational imaging fields.