Analysis Of Non-uniform Characteristics Of Diffraction Image And Research On Deblurring Algorithm

In modern optical systems,diffractive optical elements have characteristics such as light weight and flexible design,which are of great value in the field of integration or lightweight applications.They have been widely used in detection and imaging technology and other fields.However,compared with traditional optical elements,the images produced by diffractive optical elements suffer from severe degradation,and Modulation Transfer Function(MTF)has complex spatial variation characteristics.Aiming at the diffraction imaging system designed by a scientific research institute,this thesis carries out the analysis of the non-uniform characteristics of diffraction images and the research on the algorithm of deblurring.The main contents are as follows:Firstly,the non-uniform characteristics of diffraction images is studied in this thesis and an improved approach to the slanted edge method is proposed.The non-uniform characteristics of diffraction images can be characterized by MTF,and the slanted edge method is one of the main methods for measuring MTF.However,traditional methods have inaccurate edge detection under noisy conditions.A novel edge detection method is proposed which is based on improved Zernike moments.In addition,the relationship between the sampling rate and the edge angle is analyzed,and the edge diffusion function is obtained using the sampling rate that varies with the angle.Compared with ISO 12233 and ISO-cosine,the proposed method has higher edge detection accuracy,noise resistance,and improves MTF measurement accuracy.The improved slanted edge method is used to measure the MTF of each region of a radiation-shaped target,demonstrating the nonuniformity of diffraction images from the MTF perspective.Secondly,a deblurring scheme for diffraction images is studies in this thesis.Based on the classic MIMO-UNet network,corresponding improvements are made for the address the non-uniform blur of diffraction images.In the encoding process,a multi-strip attention-aware module is designed,which utilizes multi-strip feature extraction to enable the network to obtain more rich contextual information.At the same time,the attentionaware feature enhancement unit is designed using spatial and channel attention mechanisms to adaptively extract the information that needs to be focused on at each stage in the process of extracting feature information.The multi-strip attention-aware module can better learn the nonlinear features of spatially-variant blur,improving the network’s ability to handle spatially-variant blur.Finally,for the process of decoding and restoring images using the network,a dense deformable convolution residual block is designed using deformable convolution,which allows the decoding network to adaptively change the receptive field and focus on different blurry types of the diffraction images when restoring the extracted blurry features.This effectively addresses the problem of spatially-variant blur in diffraction images.The experimental results show that the proposed improvements are effective,with good performance in both subjective visual effects and objective metrics.Compared with MIMO-UNet,the proposed method achieved a PSNR of 34.62 d B on the diffraction image dataset and an improvement of 0.51 d B.Good results were also achieved on public datasets,such as a PSNR of 32.53 d B on the Go Pro dataset and an improvement of 0.9d B,verifying the effectiveness of the proposed deblurring algorithm.