Research On Adaptive Filtering Algorithms Of Multi-spectral Images Using Sparse Approximation

With the development of imaging technology,many excellent imaging methods have emerged.In the field of remote sensing imaging,multi-spectral image(MSI)can be generated with the help of physical characteristics of the sensor.Compared to traditional image acquisition systems,multi-spectral images can deliver richer information from real-world scenarios and have been proven to greatly improve the performance of a variety of computer vision tasks.With the rapid develop-ment of science and technology,multi-spectral images have been widely used in many fields such as industry,geography,agriculture,and astronomy.However,in the practical applications,multi-spectral images typically contain a certain amount of noise due to the errors of sensor acquisition system.In addition,un-der the condition of the limited radiant energy,sometimes the bandwidth is very narrow,one sensor can only capture very low energy,leads to shooting noise and thermal noise.The problems above have a negative impact on subsequent multi-spectral image processing tasks.Therefore,effective noise filtering and image enhancement can greatly facilitate the processing and analysis of multi-spectral images.Denoising has been a key and inevitable process in multi-spectral image analysis.Most of the existing multi-spectral image denoising algorithms mainly use the two-dimensional spatial information.These algorithms often ignore the structural similarity along the spectral dimension in the multi-spectral image.Even though some of the improved algorithms take into account the structural similarity among spectral bands and extend the processing domain from two-dimension to three-dimension,the details and textures of the image are still not sufficiently refined,and the information of the image is lost while removing noise.It is necessary to propose effective denoising algorithms for multi-spectral images.The paper focuses on the research and discussion of multi-spectral image quality improvement.We propose four types of multi-spectral image denoising algorithms.The main research work and innovations are summarized as the following aspects:(1)Two image denoising algorithms based on principal component analysis(PCA)are proposed.A self-similar filtering guided by principal component analysis is proposed.Non-local mean filtering is introduced to remove residual noise in the image after PCA denoising.The image denoised by the PCA twice is used as the guided image of the non-local mean filter,and the image is denoised again by the self-similarity of the image.The experimental results show that the method is superior to other competing methods in terms of fine image structures ' preservation.A color image denoising algorithm based on principal component analysis and iterative regularization is proposed.After local PCA denoising,regularization is used to supplement the image loss information,and the noise level of each similar patch group is re-estimated.Finally,the local PCA algorithm is used again on the image denoised in the previous stage to obtain the final denoised image.The experimental results show that,compared with visual observation and numerical comparison,the proposed method provides better performance than other algorithms and faithfully restores the details and texture features of the original noise-free image.(2)A multi-channel color image denoising algorithm based on hybrid thresh-old singular value decomposition(SVD)is proposed.An effective multi-channel image denoising algorithm is proposed.On thebasis of singular value decomposition,the proposed algorithm improves the tra-ditional singular value denoising algorithm,combines two excellent soft and hard thresholding algorithms,and retains the advantages of soft and hard threshold-ings.The proposed algorithm utilizes the powerful low-rank prior of non-local similar image patches.First,similar non-local cubic patches extracted from the noisy multi-spectral images are grouped into a matrix,whose columns are vector-ized patches.This matrix can be modeled by a low-rank matrix approximation.Then hybrid threshold SVD is applied to the model and a weight matrix is also introduced to balance the multi-channel characteristics of the image according to different noise levels.Alternating direction multiplier method is used to solve the estimation model,and each variable can be updated and solved by closed-form solution.Experiments on color image dataset show that the proposed algorithm is superior to existing denoising methods,both numerically and visually.(3)Two multi-spectral image denoising algorithms based on multi-scale ten-sor dictionary learning are proposed.For multi-scale tensor dictionary denoising algorithm I,tensor patches are extracted from image tensor,quadtree decomposition and zero padding are used to obtain the initial multi-scale tensor dictionary.Then,the dictionary atom is updated one by one by the multi-linear matching pursuit algorithm and the tensor decomposition to obtain the updated multi-scale tensor dictionary.The updated multi-scale tensor dictionary is then used to calculate the sparse coding of each image patch and reconstruct the image patch.Finally,all the denoised image patches are aggregated to obtain the final denoised image.For multi-scale tensor dictionary denoising algorithm II,tensor patches are extracted from the image tensor and several initial tensor dictionaries of different scales are obtained.The multi-linear matching pursuit algorithm and the tensor decomposition are used to update the dictionary atoms of each scale one by one,and the updated multi-scale tensor dictionary is obtained by zero-padding.Finally,consistent with the multi-scale tensor dictionary denoising algorithm I,the denoised image is obtained by sparse coding.Experimental results on multi-spectral images dataset show that our two methods recover more image details in visual observation than the single-scale tensor dictionary method and other methods.In addition,our methods are also superior to other existing algorithms in terms of numerical comparison.(4)A weighted Schatten p-norm minimization for multi-spectral image de-noising algorithm is proposed.For multi-spectral images contaminated by Poisson noise,a framework is pro-posed to suppress noise by low-rank matrix approximation with weighted Schatten p-norm minimization regularization.The proposed method not only considers the importance of different rank components,but also approximates the true rank of the latent low-rank matrix.This approach firstly groups similar non-local cu-bic patches extracted from the noisy multi-spectral image into a matrix,whose columns are vectorized patches.The obtained matrix can be approximated by a low-rank matrix model.Then weighted Schatten p-norm minimization is applied to the model,which shrinks different rank components with different treatments.Finally,the denoised multi-spectral image is acquired by aggregating all denoised patches.Experimental results on multi-spectral image dataset show that the pro-posed method obtains better results than state-of-the-art methods,both visually and quantitatively.