Research On Image Restoration Of Simple Lens Imaging

With the development of science and technology and the improvement of living s-tandards,the status of high-definition images in people's daily life is becoming more and more important,and optical imaging equipment is becoming more and more sophisticat-ed.Taking an SLR camera as an example,in order to ensure image quality,the lens may be a combination of a dozen different lenses.The high-precision design of the optical imaging device leads to an increase in the cost of the lens and an increase in the volumet-ric weight.In recent years,emerging computing camera technology is based on a large number of calculations,combined with digital sensors,modern optics and exciters,cre-atively utilizing clever light to get rid of the limitations of traditional cameras^[1].Based on the basic idea of?computing camera technology,this thesis attempts to improve the complex SLR lens,replace the complex SLR lens with a simple lens containing only a small number of lenses,and design the corresponding image restoration algorithm to di-rectly capture the image in the simple lens.Image blurring is performed to improve the image quality of a simple lens imaging system as much as possible.The main work and innovative points are included as following:?1?Simple lens imaging system design.In the preliminary verification stage,the simple lens group is obtained by disassembling the existing complex lens,and the feasi-bility of the simple lens imaging system combined with the computational photography idea is verified.Then CODEV software is used to design simple lens image system of one,two,three and four lenses,respectively.By analyzing the two-dimensional graph of its optical design results and the corresponding MTF?Modulation Transfer Function,MTF?curve,we explore the new design method of simple lens imaging system,and seek the best combination point of optical structure and image restoration algorithm.?2?Blind image deblurring method based on MAP.This method converts the problem of image quality improvement of simple lens imaging system directly into the problem of blind convolution image restoration.In the framework of MAP,by analyzing the PSF spatial and structural characteristics of simple lens imaging system,the method combines the blur kernel structure prior with the cross-channel prior in the optimization objective function,and uses ADMM iterative optimization algorithm to solve the objective function.Blur kernel structure prior can better reflect the PSF characteristics of simple lens imaging system,and the introduced cross-channel prior can suppress the chromatic aberration in restored image.To reduce the chromatic aberration around strong light area in the image captured by simple lens imaging system,a method based on luminance?Value?is proposed to correct the chromatic aberration,which can improve the image quality.?3?PSF estimation of simple lens imaging based on noise image pairs.To further improve the PSF estimation accuracy of simple lens imaging system,the image quality improvement is converted to non-blind convolution image restoration.Considering of imaging characteristics of simple lens imaging,the PSF is estimated with noise image pairs.Firstly,the chessboard image,black image,white image and noise image are dis-played on the computer screen.Then,the blur noise image and clear noise image pairs are obtained based on corner detection.Then,the PSF is estimated by non-blind con-volution method.Then,the non-blind convolution image is restored based on estimated PSF.Compared with the common blind convolution PSF estimation method,the proposed method can effectively improve the PSF estimation accuracy and then improve the image restoration quality.?4?Fast non-blind image deblurring in frequency domain based on matrix decompo-sition.Traditional frequency-domain non-blind convolution image restoration method-s can not meet the real-time requirement when processing large images.This method attempts to accelerate its algorithm by using frequency-domain matrix decomposition method.Firstly,the frequency-domain matrix of the blurred image is divided into a series of linear combinations of bases.The existing non-blind convolution algorithm is used to process each base and preserve the result.For the new blur image,the matrix is also divided into linear combinations of bases,and the corresponding linear combinational co-efficients are obtained.Simple multiplication and addition operations can be performed by the linear combination coefficients and the preserved bases.The restored clear image is obtained.Experimental results demonstrate that the proposed method achieves almost the same deblurring results while the computing speed can improve by 40%.?5?End-to-End image deblurring method based on Encoder-Decoder network.The semi-simulated deep learning dataset of simple lens imaging can effectively avoid the influence of the error during the image acquisition process on the accuracy of dataset.The encoder part gradually generates the feature diagrams of each level,the decoder part is further analyzed and processed based on the feature diagrams of the input image and the encoding part,at the same time,the residual learning is combined with the encoder-decoder network,and the encoder unit is used to extract and process the image features at each resolution level.It is beneficial to improve the learning ability of the network and promote the gradient propagation in the training process.The L2 norm loss function is used to supervise the training process of the network.Experimental results demonstrate that the proposed method achieves better deblurring result on the simulated dataset,and achieve almost the same deblurring result for real images captured by simple lens imaging.