CN105654425A - Single-image super-resolution reconstruction method applied to medical X-ray image - Google Patents

Abstract

The invention discloses a single-image super-resolution reconstruction method applied to a medical X-ray image. The method comprises a dictionary training process and a super-resolution reconstruction process. The dictionary training process is characterized by through using similarities of details and structures of same modal medical images, collecting similar high-quality X-ray images as samples and establishing a database of a high-low resolution image block. The super-resolution reconstruction process is characterized by carrying out blocking reconstruction on an input low-resolution image to generate the image block and then reconstructing a whole high resolution image. By using a super-resolution reconstruction algorithm based on sparse representation, the medical X-ray image can be amplified on an aspect of a space scale, and simultaneously PSNR and SSIM values are increased compared to the PSNR and SSIM values by using other algorithms, at the same time, noise robustness is increased, a calculating complexity is reduced and calculating consumption time is reduced. Impression and a general method measurement index of the reconstruction image are increased.

Description

A kind of single image super resolution ratio reconstruction method being applied to medical science x-ray image

Technical field

The present invention relates to Image Reconstruction Technology field, more particularly a kind of be applied to medical science x-ray image based on rarefaction representation single image super resolution ratio reconstruction method.

Background technology

Image procossing is as one of important technology being widely used in the information processing technology, and development in recent years is rapid. As the important indicator weighing image processing techniques, the resolving power of image superior quality has growing demand in all conglomeraties. Such as in medical image field, medical science image has become the important evidence that clinician diagnoses the state of an illness. High-resolution clear X-ray is photographed, and is conducive to the diagnosis to some disease. Along with the raising of medical requirement, the resolving power of x-ray image and sharpness be it is also proposed higher requirement by people. Meanwhile, in tele-medicine field, the storage of medical science image and video and transmission are also needed storage and the display of high resolution image badly.

But, it being limited to acquisition technique level and hardware cost, the image that we obtain usually can not meet higher resolution requirement, the medical image field related to especially herein, and the meeting on the low side of image resolution rate causes doctor to the false judgment of disease. In addition, in the acquisition process of medical image, the dosage of radioactive substance and apparatus operating time have strict specification limits, and imaging process is more vulnerable to external interference, cause the medical science image of generation to be mixed into noise, reduce picture quality. Therefore, in order to the application needs of satisfied reality, we improve a kind of method of image procossing that can be used for carrying on the software of high image resolution, namely based on the super-resolution image reconstruction method of rarefaction representation, it is to increase the effect simultaneously reaching reduction noise of image resolution rate.

Image super-resolution rebuilding is a kind of image processing method being reconstructed high-resolution image by the image of a width or a few width low resolution, in order to break through the restriction of sensor low resolution. Its application value is at medical field and obviously. Obtain X-ray photography and the nuclear magnetic resonance of high-resolution, be conducive to the diagnosis to some disease, also contribute to effective implementation of remote operation and diagnosis. The key concept of algorithm is proposed in the sixties in 20th century by Harris and Goodman the earliest, but they do not obtain concrete results in practical application area.Until early 1980s Tsai and Huang utilizes first does not fall several down-sampled low-resolution frames clear, translation, super-resolution rebuilding has gone out single width high-definition picture. Current image super-resolution method is mainly divided into based on many frame image sequence with based on the Enhancement Method of single two field picture. Concrete grammar is mainly divided into the method based on interpolation, the method based on reconstruction and the method based on study. Super-resolution method based on single two field picture comprises the edge enhancing method without learning sample and has the method based on study of learning sample. Wherein based on study super-resolution method be research focus in recent years, it utilizes training sample set to set up high-frequency information model by machine learning, and then the high-frequency information of prediction low resolution test sample book, obtaining high-definition picture, adding of priori makes reconstruction image comprise abundanter high-frequency information.

Super-resolution rebuilding technology is applied to the concern that this problem of medical science image also result in a lot of researchist, in PET image, MRI image, ultrasonic wave image etc., achieve effective achievement at present, but how to improve the medical science image being seriously subject to noise jamming, while improving spatial resolution, effectively reduce noise, it is still a problem being rich in challenge.

There is the requirement to medical science image high-resolutions such as X light, CT in clinical diagnose and research field, and it being limited to medical image acquisition device, technology and HardwareUpgring cost, we need the Medical Image Processing technology improving this field in software: resolving power and the sharpness improving image on the one hand; Improve image processing speed on the other hand, promote the efficiency that the high resolving power of medical science image stores and transmits.

For addressing this problem, the present invention proposes a kind of new dictionary learning method,

Summary of the invention

Based on above-mentioned prior art, the present invention proposes a kind of single image super resolution ratio reconstruction method being applied to medical science x-ray image, it is proposed to new dictionary learning method, utilize PCA principal component analytical method simultaneous training low resolution and high resolving power dictionary block. Then for each low-resolution image block of input, low-resolution dictionary is found the sub-dictionary the most similar to it, obtain the sub-dictionary of high resolving power corresponding with it, solve super-resolution rate problem with the graceful linear iterative algorithm of Burger, generate high-definition picture block.

A kind of single image super resolution ratio reconstruction method being applied to medical science x-ray image of the present invention, the method comprises the following steps:

Step 1, dictionary training process: first, high-definition picture is as follows with a set expression that overlapping image block is formed:

$I_{H-input}={\mathrm{Blk}}_i^H,i=1,2,...,N.---\left(1\right)$

Wherein,Being a size isImage block, N is from image I_H-inputThe number of the image block of middle generation;

Then extract operator with Hi-pass filter as feature, extract eachThe high frequency composition of image block, obtains feature matrix I_f;

By feature matrix I_fK class I is obtained after carrying out cluster₁,I₂,��,I_k, then the center of each class is designated as M₁,M₂,��,M_kSo that each feature matrix to its nearest center all side apart from minimum;

The covariance matrix of each correspondence is designated as �� k, PCA principal component analytical method is applied to each �� k, obtains orthogonal transform matrix P_k, extract r wherein relatively important proper vector and form dictionary matrix D_k, D_k=[p₁,p₂,��,p_r], it is determined that the optimum solution r of r₀:

$r_0=\arg \underset{r}{\min }\left(\right||I_k-D_k{\mathrm{\α}}_k|{|}_2^2+\mathrm{\λ}\left|\right|{\mathrm{\α}}_k\left|{|}_1\right).---\left(2\right)$

By the optimum solution r solved₀And then obtain K sub-dictionaryNamely the sub-dictionary of K high resolving power that K class is corresponding is obtained;

To all high resolving power center M₁,M₂,��,M_kCarry out fuzzy and lower sampling so that each high resolving power center correspond to a unique low resolution center, obtain the sub-dictionary of low resolution corresponding to the sub-dictionary of high resolving power with reason;

Step 2, super-resolution rebuilding process: the low resolution x-ray image of test input is divided into image block; The high frequency component of each low-resolution image block is extracted with Hi-pass filter, and by itself and K low resolution centerCompare, find the class matched most, it is determined that the sub-dictionary of low resolution, so that it is determined that the sub-dictionary of the high resolving power of correspondence;

With the low-resolution image block x of input_lEstimate the high-definition picture block x exported_h, it is necessary to solve following sparse coding problem and obtain rarefaction representation coefficient:

$\mathrm{\α}=\arg \underset{\mathrm{\α}}{\min }\left(\right||I_l-\mathrm{\φ}\mathrm{\α}|{|}_2^2+\mathrm{\λ}\left|\right|\mathrm{\α}\left|{|}_1\right).---\left(3\right)$

In formula, ��=MBD, D are lower sample operator, and M is motion operator, and B is fuzzy operator.

Applying linear Bregman iterative algorithm and solve the rarefaction representation vector �� that formula (3) obtains low-resolution image, corresponding high-definition picture block can be represented as:

x_h=D_new��(5)

Estimate whole high-definition picture blocks, reconstruct high-definition picture.

Experimental comparison results shows, while medical science x-ray image can be carried out the amplification on space scale by the super-resolution rebuilding algorithm based on rarefaction representation that the present invention proposes, relatively PSNR and the SSIM value of other algorithms promotes to some extent, enhance noise robustness simultaneously, reduce computation complexity, reduce the calculation consumption time; And achieve the raising rebuilding image in perception and general method measurement index.

Accompanying drawing explanation

Fig. 1 is the overall flow figure of a kind of single image super resolution ratio reconstruction method super resolution ratio reconstruction method being applied to medical science x-ray image of the present invention;

Fig. 2 is the high-resolution dictionary center that obtains of training and the center schematic diagram of corresponding low-resolution dictionary: the center carrying out low resolution that dictionary training obtains and high resolving power class with the collection of training image shown in Fig. 5, and low resolution center obtains after sampling and Fuzzy Processing by under high resolving power center;

Fig. 3 is the sub-dictionary schematic diagram learning to obtain, and learns the sub-dictionary schematic diagram obtained, and left side represents high resolution class center, and right side represents corresponding low resolution class center;

Fig. 4 is experimental result example schematic diagram: (4-1) from left to right under be followed successively by: the low-resolution image that Gaussian noise grade is; Bicubic interpolation (PSNR=33.25, SSIM=0.81); NEDI (PSNR=32.88, SSIM=0.90); ScSR (PSNR=33.80, SSIM=0.88); This paper method (PSNR=36.37, SSIM=0.92); Original high-resolution image; (4-2) from left to right under be followed successively by: Gaussian noise grade is the low-resolution image of ��=5; Bicubic interpolation (PSNR=36.27, SSIM=0.69); NEDI (PSNR=36.15, SSIM=0.91); The ScSR (PSNR=37.45, SSIM=0.77) of ��=0.2; This paper method (PSNR=38.97, SSIM=0.94); Original high-resolution image;

The training image collection exemplary plot used when Fig. 5 is experiment, (5-1) is for building the standard map picture collection lung front of dictionary; (5-2) for building the standard map picture collection lung side of dictionary.

Embodiment

The present invention represents on the basis of ultimate principle and image super-resolution technological frame at image sparse, for dictionary training part, it is proposed to a kind of method for reconstructing for x-ray image newly. Dictionary learning process utilizes PCA principal component analytical method simultaneous training high-low resolution dictionary block, reduce and rebuild the high noise phenomenon that may exist of image, it is to increase algorithm efficiency;At process of reconstruction resume based on falling the consideration made an uproar, linear Bregman iterative algorithm is adopted to rebuild image block.

Below in conjunction with the drawings and the specific embodiments, it is described in further detail the technical scheme of the present invention.

As shown in Figure 1, overall technical architecture is as follows:

Step 1, dictionary training process

Utilize mono-modality medical image in details and structural similarity, gather similar high-quality x-ray image as sample, set up the database of height-low-resolution image block. In order to save computation complexity when image block is carried out sparse coding, improve algorithm to the robustness of noise simultaneously, utilize clustering method to obtain K sub-dictionary, and adopt PCA principal component analytical method to set up dictionary. Process is as follows:

First, (1) that high-definition picture is as follows with a set expression that overlapping image block is formed:

$I_{H-input}={\mathrm{Blk}}_i^H,i=1,2,...,N.---\left(1\right)$

Wherein,Being a size isImage block, N is from image I_H-inputThe number of the image block of middle generation.

(2) then extract operator with Hi-pass filter as feature, extract eachThe high frequency composition of image block, obtains feature matrix I_f��

(3) K class I is obtained after feature matrix being carried out cluster₁,I₂,��,I_k, then the center of each class is designated as M₁,M₂,��,M_kSo that each feature matrix to its nearest center all side apart from minimum.

(4) covariance matrix of each correspondence is designated as �� k, PCA principal component analytical method is applied to each �� k, obtains orthogonal transform matrix P_k. Extract r wherein more important proper vector and form dictionary matrix D_k, D_k=[p₁,p₂,��,p_r]. Determine the optimum solution of r by formula (2), optimum r is designated as r₀:

$r_0=\arg \underset{r}{\min }\left(\right||I_k-D_k{\mathrm{\α}}_k|{|}_2^2+\mathrm{\λ}\left|\right|{\mathrm{\α}}_k\left|{|}_1\right).---\left(2\right)$

(5) by the r solved₀And then obtain K sub-dictionaryNamely the sub-dictionary of K high resolving power that K class is corresponding is obtained.

(6) for avoiding the image block according to input to find not mating in dictionary process, double word allusion quotation is adopted to carry out super-resolution rebuilding herein. Herein to all high resolving power center M₁,M₂,��,M_kCarry out fuzzy and lower sampling so that each high resolving power center correspond to a unique low resolution center, as shown in Figure 3. The sub-dictionary of low resolution corresponding to the sub-dictionary of high resolving power is obtained with reason.

The center of high-resolution dictionary center and corresponding low-resolution dictionary is as shown in Figure 2.

Step 2, super-resolution rebuilding process

When the low-resolution image of input being carried out piecemeal and rebuilds, rebuild image to the robustness of noise for improving, apply linear Bregman iterative algorithm and solve and obtain corresponding sub-dictionary and each is as the rarefaction representation coefficient of block based on this training dictionary. Thus synthetic image block, and reconstruct whole width high-definition picture further. Process is as follows:

(1) the low resolution x-ray image of test input is divided into image block by experiment.

(2) extract the high frequency component of each low-resolution image block with Hi-pass filter, and itself and a low resolution center are compared, find the class matched most, it is determined that the sub-dictionary of low resolution, so that it is determined that the sub-dictionary of the high resolving power of correspondence.

(3) the high-definition picture block exported is estimated with the low-resolution image block of input, it is necessary to solve following sparse coding problem and obtain rarefaction representation coefficient:

$\mathrm{\α}=\arg \underset{\mathrm{\α}}{min}\left(\right||I_l-\mathrm{\φ}\mathrm{\α}|{|}_2^2+\mathrm{\λ}\left|\right|\mathrm{\α}\left|{|}_1\right).---\left(3\right)$

In formula, ��=MBD, D are lower sample operator, and M is motion operator, and B is fuzzy operator.

Obtaining rarefaction representation vector �� to solve formula (3), the present invention applies linear Bregman iterative algorithm.

The present invention use one group of size be 885 �� 885 lung front and lung's side high resolving power x-ray image as training image, the size of high-definition picture block and low-resolution image block is respectively 7 �� 7 and 3 �� 3.

In all experiments, low-resolution image generates with corresponding high-definition picture, is specially three steps: first, is 1.6 with standard deviation, size be 7 �� 7 gaussian filtering operator training image is carried out fuzzy; Then with decimation factor s, training image is carried out lower sampling; Finally the image degenerated is added the white Gaussian noise that standard deviation is ��.

For all methods, the magnification of Setup Experiments is 3. During experiment, the �� in formula is set to 0.8, is set to 0.7, is set to 0.01. The maximum iteration time of algorithm is set to 999, exceedes this number of times and then terminates this section of program.

Use the objective quality of two quality metric method evaluate image method for reconstructing herein, one is peak value signal to noise ratio (PeakSignaltoNoiseRatio, PSNR), another kind is structural similarity (StructureSimilarity, SSIM).

Experimental result is as shown in Table 1, learn by observing, method in this paper is all superior to bicubic interpolation (BicubicInterpolation on PSNR and SSIM, BI), slightly it is superior to the super resolution ratio reconstruction method (SparseCodingSRmethod of sparse dictionary coding, and new edge-directed interpolation (NewEdge-DirectedInterpolation, NEDI) ScSR).

Table one, inventive algorithm and part super-resolution algorithms rebuild effectiveness comparison (scale-up factor is set to 3)

Magnification s=3, as shown in Figure 4, the lower right corner is area-of-interest (DesiredRegionofInterest, DROI) for the lung front of noise grade ��=5 and side image. Meanwhile, the working time of the dictionary training needed for method that the present invention proposes only needs 209 seconds, is only only training 512 to 1/10th of the ScSR method of image block double word allusion quotation. This has benefited from PCA principal component analytical method and the sub-dictionary method superiority in application.

Claims (1)

1. one kind is applied to the single image super resolution ratio reconstruction method of medical science x-ray image, it is characterised in that, the method comprises the following steps:

Step (1), dictionary training process: first, high-definition picture is as follows with a set expression that overlapping image block is formed:

$I_{H-input}={\mathrm{Blk}}_i^H,i=1,2,...,N.---\left(1\right)$

Wherein,Being a size isImage block, N is from image I_H-inputThe number of the image block of middle generation;

Then extract operator with Hi-pass filter as feature, extract eachThe high frequency composition of image block, obtains feature matrix I_f;

By feature matrix I_fK class I is obtained after carrying out cluster₁,I₂,��,I_k, then the center of each class is designated as M₁,M₂,��,M_kSo that each feature matrix to its nearest center all side apart from minimum;

The covariance matrix of each correspondence is designated as ��^k, PCA principal component analytical method is applied to each ��^k, obtain orthogonal transform matrix P_k, extract r wherein relatively important proper vector and form dictionary matrix D_k, D_k=[p₁,p₂,��,p_r], it is determined that the optimum solution r of r₀:

$r_0=\arg \underset{r}{min}\left(\right||I_k-D_k{\mathrm{\α}}_k|{|}_2^2+\mathrm{\λ}\left|\right|{\mathrm{\α}}_k\left|{|}_1\right).---\left(2\right)$

By the optimum solution r solved₀And then obtain K sub-dictionaryNamely the sub-dictionary of K high resolving power that K class is corresponding is obtained;

To all high resolving power center M₁,M₂,��,M_kCarry out fuzzy and lower sampling so that each high resolving power center correspond to a unique low resolution center, obtain the sub-dictionary of low resolution corresponding to the sub-dictionary of high resolving power with reason;

Step (2), super-resolution rebuilding process: the low resolution x-ray image of test input is divided into image block; The high frequency component of each low-resolution image block is extracted with Hi-pass filter, and by itself and K low resolution centerCompare, find the class matched most, it is determined that the sub-dictionary of low resolution, so that it is determined that the sub-dictionary of the high resolving power of correspondence;

(3) with the low-resolution image block x of input_lEstimate the high-definition picture block x exported_h, it is necessary to solve following sparse coding problem and obtain rarefaction representation coefficient:

$\mathrm{\α}=\arg \underset{\mathrm{\α}}{min}\left(\right||I_l-\mathrm{\φ}\mathrm{\α}|{|}_2^2+\mathrm{\λ}\left|\right|\mathrm{\α}\left|{|}_1\right).---\left(3\right)$

In formula, ��=MBD, D are lower sample operator, and M is motion operator, and B is fuzzy operator.

Applying linear Bregman iterative algorithm and solve the rarefaction representation vector �� that formula (3) obtains low-resolution image, corresponding high-definition picture block can be represented as:

x_h=D_new��(5)

Estimate whole high-definition picture blocks, reconstruct high-definition picture.