Nonsubsampled Contourlet Transform (NSCT)
Similar to the wavelet transform, the source image is transformed after some processing on the coefficient, and then invert back to the target image.
NSCT is a new translation invariant, multi – scale, multi – directional fast transform.
NSCT is a transformation based on the Nonsubsampled pyramid (NSP) and the Nonsubsampled Directional filter (NSDFB). Firstly, the input image is decomposed into high-pass and low-pass parts by NSP, and then the high-frequency subband is decomposed into multiple directional subbands by NSDFB, and the low-frequency part continues to be decomposed as above.
- Nonsubsampled Pyramid (NSP) :
Nonsubsampled Pyramid(NSP) and Contourlet’s Laplacian Pyramid(LP) multiscale analysis have different characteristics. The image is multiscale decomposed by Nonsubsampled Pyramid (NSP), which removes both up-sampling and down-sampling, reducing the sampling distortion in the filter and achieving translation invariance. NSP is an NSCT multi-scale analysis with translational invariant filtering structure, which can obtain the same multi-scale analysis characteristics as LP decomposition. FIG. 2.4(a) is divided into three scales.
- Nonsubsampled Directional filter (NSDFB)
The Nonsubsampled Directional filter (NSDFB) is a two-channel filter that synthesizes the coefficients of NSCT from singularities distributed in the same direction. The directional filter (DFB) was proposed by Bamberger and Smith. Through the decomposition of an L-layer tree structure, the signal is effectively divided into sub-bands, and the band is segmented into a wedge. Nonsubsampled DFB (NSDFB) is non-sampled, reducing sampled distortion in the filter and achieving translation invariance. In addition, the size of the directional subgraph at each scale is the same as that of the original graph, and the Contourlet transform is that the sum of all subbands is equal to the original graph. NSCT has more detail preserved and the transformation coefficients are redundant. Below is the image frequency domain segmentation map at three scales, where the number of direction subbands at each scale increases by 2 times, and the number of direction subbands at 1, 2 and 3 scales is 2, 4 and 8 respectively. Paper 125: NSCT – Nonsubsampled contourlet transform program (invariance scale) – Jason. Hevey – blog garden www.cnblogs.com/molakejin/p…
NSCT: Nonsubmediated-Contourlet Transform algorithm and MATLAB Implementation _ Computer Vision CV, Image Processing, Image Fusion _ SOaringlee_Fighting Blog.csdn.net/soaringlee_…
Nonsubsampled Contourlet transform (Nonsubsampled Contourlet) Transform (NSCT) is composed of a non-sampling pyramid structure and a non-sampling direction filter bank, which has excellent characteristics such as translation invariance and high redundancy that Contourlet transform does not have, and can overcome the pseudo-Gibbs phenomenon. The image is decomposed into multi-scale and multi-direction details after non-subsampled Contourlet transformation. These details represent the features of different directions in different frequency bands of the image, which simplifies the relationship between coefficients.
Study Paper 1:
Combining NSCT with Interpolation for Image Super-resolution reconstruction – Doug Baba www.doc88.com/p-549641892…
Study Paper 2:
Fingerprint Image Super-resolution reconstruction based on learning NSCT.
Study Paper 3:
A new kind of NSCT super-resolution image restoration technology. PDF max.book118.com/html/2015/0…
Study Paper 4:
Combined with the feature of much of single image super-resolution reconstruction algorithm – Harbin industrial university scientific research academic – Free one’s deceased father grind school.freekaoyan.com/heilongjian…
In order to improve the quality of directly captured images, a super-resolution reconstruction algorithm combining Gabor transform and NSCT was proposed to solve the problem that only horizontal and vertical information could be extracted from gradient features and details could not be extracted from non-subsampled contours transform (NSCT). This algorithm makes full use of the complementarity of Gabor transform and NSCT. According to the characteristics of the input image blocks, Gabor transform is used to extract texture features and NSCT to extract contour features, and then sparse model is used to reconstruct them, and finally merged into a high-resolution image. Since the input image is more or less fuzzy, the regular term of defuzzy is added in the reconstruction process to eliminate the influence of input fuzzy. Experimental results show that the super-resolution effect of combining the two features can recover more details than that of a single feature, and defuzzy regular terms also play a certain role. Compared with the kernel ridge regression proposed by Kim and sparse representation (SCSR) proposed by Yang, the visual effect of the proposed method is more clear, and the PSNR value is improved by nearly 2dB on average objectively, indicating that the proposed algorithm can effectively improve the image quality.
Variance classification: For flat blocks, texture information is restored, and Gabor filter is used to extract texture features of small blocks. For non-flat blocks, the contour information is recovered, and NSCT is used to extract the contour features of small blocks.
NSCT not only has good multi-directional and multi-scale properties, but also has translation invariance, which can effectively extract contour features of images. NSCT consists of non-subsampled pyramid (NSP) decomposition and non-subsampled direction filter bank (NSDFB) decomposition.
Taking two-layer NSP decomposition and two-layer NSDFB decomposition as examples, the image decomposition process is described. Firstly, the image is decomposed by NSP to obtain a high frequency subband and a low frequency subband. Then, NSDFB is used to decompose the high frequency subband. If the direction decomposition coefficient is M, 2m high frequency subbands are obtained. The next NSP decomposition will decompose the low-frequency image obtained by the previous NSP decomposition. Therefore, after two-layer NSP decomposition and two-layer NSDFB decomposition, one low-frequency sub-band and eight high-frequency sub-bands can be obtained. Compared with gradient features, NSCT can obtain more scale and direction information, and has a stronger ability to describe the image.
%NSCT_imagefusion% % achieve infrared and visible imagefusion; % clear all; close all; clc; % initialization ends % % timing starts % tic; path(path,'nsct_toolbox'); path(path,'ziji'); Kk = 1.5; % % % % % % % % % % % input image % % % % % % % % % % % [input_image_TV_int_RGB input_image_TV_int_MAP] = imread (' infrared. JPG); % input_image_TV_int_RGB = uint8(256* ind2RGB (input_image_TV_int_GIF, input_image_TV_int_MAP)); % convert GIF data to RGB data % [input_image_IR_int_RGB, input_image_IR_int_MAP] = imread(' visible light.jpg'); % Input GIF format infrared image % % input_image_IR_int_RGB = uint8(256* ind2RGB (input_image_IR_int_GIF, input_image_IR_int_MAP)); % to GIF format data is converted to RGB format data % % % % % % % % % % % % % % % end of the input image % % % % % % % % % % % % % % [Ny, Nx] = size (input_image_TV_int_RGB); % % % % % % % % % % % % image preprocessing % % % % % % % % % % % % % % of the visible light image format conversion for IHS format input_image_TV_int_IHS = rgb2ihs (input_image_TV_int_RGB); input_image_TV_IHS = double(input_image_TV_int_IHS); % Convert the data type of the input image to a double data type % input_image_TV_I = 256*input_image_TV_IHS(:,:,3); % Extract the I component of the input image for NSCT decomposition % % Convert the format of the infrared image to gray format % input_image_IR_int = RGB2Gray (input_image_IR_int_RGB); input_image_IR= double(input_image_IR_int); % Convert the data type of the input image to a double data type % input_image_TV_S = input_image_TV_IHS(:,:,2); input_image_TV_H = input_image_TV_IHS(:,:,1); figure; Subplot (1, 2, 1); imshow(uint8(input_image_TV_I)); Title ('Original TV image I') % displays the I component of the input visible image % subplot(1,2,2); imshow(uint8(input_image_IR)); Title (' the Original IR image Gray ') display input infrared image grayscale image % % % % % % % % % % % % % % % % % % image preprocessing end % % % % % % % % % % % % % % % % % pyramid said Nsc parameter % = ceil(log2(min(Ny,Nx)) - 7); % Number of decomposition scales (adaptive to image size)% Nor = 8; Each level of decomposition in the direction of the % % % % % % % % % % % % % % % % % % end of the initialization parameter % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % NSCT subband decomposition % % % % % % % % % % % % % % % % initialization NSCT subband decomposition pfilter parameter % = 'maxflat' ; % pyramid filter % dfilter = 'dmaxFlat7'; % nlevels=zeros(1,Nsc+1); for i=1:Nsc+1 nlevels(i)=log2(Nor); % %NSCT % coeffs_TV_int = NSCtDEC (input_image_TV_I, nlevels, dfilter, pfilter); % Decomposed visible image % coeffs_IR_int = nSCTDEC (input_image_IR, nlevels, dfilter, pfilter); % % % NSCT decomposition infrared image decomposition end % % % % % % % % % % % % % % % % % % % NSCT subband decomposition end % % % % % % % % % % % % % % % % % % % imwrite(uint8(coeffs_IR_int{1}),'lowfrequent.bmp','bmp'); % % % % % % % % % % % image fusion % % % % % % % % % % % % % % % % % % % % % % % % % image fusion end % % % % % % % % % % % % % % % % % % % % % % % % % % % % % NSCT subband refactoring % % % % % % % % % % % % % % % out_image_end_I = nsctrec( coeffs_rec, dfilter, pfilter ) ; % Reconstructed gray image % figure; imshow(uint8(out_image_end_I)); Title (' fused image I'); out_image_end=zeros(Ny,Nx,3); % initialize color image output matrix % out_image_end=cat(3,input_image_TV_H,input_image_TV_S,out_image_end_I/256); out_image_end_RGB=256*ihs2rgb(double(out_image_end)); % Convert the output color image to RGB format % figure; imshow(uint8(out_image_end_RGB)); % imwrite(uint8(out_image_end_RGB),' swt22.bmp ',' BMP '); toc;Copy the code
