LMMSE-based low complexity image denoising in non-uniform directional filter bank
This dissertation is about creating a non-redundant, effective and low-complexity denoising method. Denoising an image involves removing noise from an image to keep the original elements of the image and remove unwanted additions. Transform-based denoising depends on the transform used in the denoi...
Saved in:
| Format: | Thesis |
|---|---|
| Language: | English |
| Subjects: | |
| Online Access: | http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/77165/1/Page%201-24.pdf http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/77165/2/Full%20text.pdf http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/77165/4/Yahya%20Naji%20Saleh.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This dissertation is about creating a non-redundant, effective and low-complexity denoising method. Denoising an image involves removing noise from an image to keep the original elements of the image and remove unwanted additions. Transform-based
denoising depends on the transform used in the denoising method. Recent works focus on improving the performance of the denoising, ignoring the complexity. This work studies the complexity of well-known transforms that capture texture in images
sufficiently, known as Contourlet transform (CT) and non-uniform directional filter bank (NUDFB). Complex wavelet transform (CWT), CT and non-subsampled Contourlet transform are examples of the base of denoising methods in the majority of current work. All these transforms perform well but they are complex and involve redundancy. Noise applied to images in this work is Gaussian noise and all images used are greyscale; the Linear Minimum Mean Square Error (LMMSE) method was used for denoising, with NUDFB as the base transform. The method decomposes the image using NUDFB then recognizes the coefficients to the step of denoising. LMMSE compares the coefficients in different resolutions to predict and remove noise. For measuring work accuracy, Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are used. The resulting
images from the proposed denoising method show improvement. The PSNR of the proposed method is higher than thresholding using NUDFB, about 1db. When comparing the proposed method with wavelet transform thresholding and CT, it has higher values than in CT and WT, especially in a high noise ratio. In images that contain directional information, such as fingerprint images, the proposed method has the highest SSIM. The proposed denoising method creates a way of denoising images using fewer requirements because of the low complexity. LMMSE also perform better than thresholding method. |
|---|