Research On Nonlocal Transform Domain Image Denoising And Enhancement And Their Performance Evaluation

Image denoising is a basic topic in image processing research. The existing image denoising methods are classified into local methods and nonlocal methods, the nonlocal means (NLM) method as a brand-new image denoising strategy was proposed in recent years. The recently developed block-matching 3D collaborative filtering (BM3D) effectively combined local transform method with the nonlocal idea, it is recognized as the best image denoising method at present. This paper profoundly analyzes some drawbacks of BM3D method and proposes several improved BM3D algorithms. The proposed algorithms achieve better image denoising results than the original BM3D method. Image detail enhancement is another important research topic in image processing. A premise of transform domain image detail enhancement is that weak detail information can be effectively represented in transform domain, amplify weak detail coefficients to achieve the image detail enhancement. As the currently best linear singular representation method, nonsubsampled contourlet transform (NSCT) can effectively represent the edge or texture image detail information, but the local transform methods achieve image detail representation by the convolution operation between certain convolution kernels and images, the inverse transform after the weak detail coefficients enhancement usually introduce serious halo phenomenon around the image edges. In this paper, a nonlocal linear singular representation method is proposed, applying it to image enhancement the ideal image enhancement results can be achieved, and few halo artifacts are introduced. For image denoising and enhancement performance evaluation, this paper proposes an image watermark based image denoising and enhancement performance evaluation method. The main research contents and progress are given below.BM3D firstly implements the block-matching to find some similar image blocks and stacks them to form a 3D array, a separable 3D transform is then implemented on the 3D array, the hard thresholding shrinkage operation on the transformed coefficients is used to achieve image denoising. BM3D can better preserve the image details then NLM, but the 2D transformation on each block is still a local one, so some artifacts will be introduced. Especially when noise levels are relatively higher, not only the artifacts are introduced but also the denoising performance will sharply drop. In this paper, how to introduce less artifacts and how to effectively solve the problem of denoising performance sharp drop when the noise levels are relatively higher are both profoundly studied. Through the analysis of the number of the matched image blocks, the size of the blocks, the influence of pre-filtering on each block before block-matching, this paper proposes a better approach than the original BM3D for resolving the aforementioned problems. The proposed method has better image denoising performance, especially less artifacts are introduced by using the proposed method than original BM3D.The powerful denoising performance of BM3D method is rooted in the enhanced sparsity representation, the enhanced sparsity is from the high similarity among the blocks in each image blocks group and the separable 3D transform on each 3D array. But the BM3D overly emphasizes the sparsity. In the practical implementation of BM3D, a separable 3D transform is implemented on the 3D array, i.e.,2D transform is implemented on each image block, and then an interblock third dimensional transform is implemented. This paper proposes to change the 3D transform of BM3D to two steps interblock 1D transform, which increases the sparsity among the blocks, but weakens the sparsity inner each block, so it can less introduce artifacts. Because the two steps 1D interblock transform uses less image blocks, the preservation of image details is also better than the original BM3D method. The proposed method in this paper is called block-matching ID and 3D collaborative filtering (BM1-3D).From the human visual perception, when an image is added the same level noise in every area the smooth area looks noisier than the image edges and textures. Based on this fact, this paper proposes a morphological component based size adaptive block-matching transform domain filtering image denoising algorithm. According to the energy the alternating current (AC) coefficients of discrete cosine transform (DCT) on image blocks, all the image blocks are divided into smooth, contour or texture three kinds of morphological components. In order to improve the denoising performance, different block sizes for different morphological components are used to implement block-matching 1D-3D transform domain filtering. The experimental results show that both objective evaluation and subjective visual quality of the denoising results by the proposed method are consistently better than the BM3D method.Contours and textures are the most important information in natural images and they usually present linear singularity, transform based image enhancement methods usually represent image details in advance, the image enhancement can be achieved by amplifying weak image detail coefficients. In view of the classic orthogonal wavelet transform can only well represent point singularity, in order to better represent the linear singularity, NSCT has been proposed and applied to the image enhancement. But no matter the orthogonal wavelets or beyond wavelets always decompose an image by the convolution operation between the convolution kernels and the image, especially when some larger support filters are used, some strong halo artifacts will be introduced after the inverse transform. In this paper, a nonlocal linear singularity representation method is proposed which is called block-extracting and Haar transform (BEH). A relatively small neighborhood's center is considered as a reference point, K image blocks are extracted whose upper left coordinates are around the reference point, because all the extracted image blocks by this strategy have similar smooth background, but the image details locations in these blocks are usually different with each other, when an interblock Haar transform is implemented on the image block group formed by all the extracted blocks, the linear singularity can be effectively represented. This paper uses this nonlocal linear singularity method to image enhancement to verify its effectiveness, an image enhancement algorithm is also proposed based on BEH. The proposed image enhancement algorithm introduces less halo artifacts but achieves ideal image details enhancement. Experimental results demonstrate that both objective evaluation and subjective visual quality of the enhanced images by the proposed method are significantly better than some existing image enhancement methods.Image watermark is an effective means of digital image copyright protection, this paper proposes a semisubsampled wavelet transform, and the watermark is embedded in the low frequency sub-band of transform in order to improve the transparency and robustness of the watermark. The watermark is in essence a kind of additive noise, this paper proposes an image watermark based image denoising and enhancement performance evaluation method, evaluate image denoising and enhancement performance according to the bit error rate (BER) of the extracted watermark from the denoised or enhanced watermarked image. For the geometric attack, this paper proposes two watermark resynchronization algorithms for rotation and shearing attacks respectivly. The experimental results show that the proposed watermark based image denoising and enhancement performance evaluation method is feasible and effective; the proposed watermark resynchronization algorithm for geometric attacks can effectively resynchronize the desynchronized watermark by rotation attack and shearing attack.