gnixihsil / image-denoising-2

A project testing and comparing different image denoising methods.

Image-Denoising

This project applies different image de-noising methods on natural and synthetic images that are distorted by level 5 AWGN (Additive white Gaussian noise), and conducts a comparative evaluation on the results of different methods in terms of quantity and quality.

Program Usage

1. Open the “Test.m” script in the “code” Folder in Matlab.
2. Run the “Test.m” script.

Parameter Settings

• For wavelet methods, both hard and soft filters are tested. We choose Biorthogonal 3.5 as the wavelet filter when comparing wavelet with other method. The DWT level is set as 3.
• For BM3D, the sigma value is set as 25.
• For spatial filters, the kernel size is set as 5x5.
• For Contra Harmonic Mean Filters, the positive and negative Q parameters are set as 1.5 and -1.5 respectively.

Evaluation Metrics

• For quantitative analysis, the evaluation metrics include the well-known PSNR as well as PSNR-HVS-M suggested by Nikolay Ponomarenko, et al (On between-coefficient contrast masking of DCT basis functions, CD-ROM Proceedings of the Third International Workshop on Video Processing and Quality Metrics for Consumer Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p).

• For qualitative analysis, we observe the de-noised images visually to make subjective evaluations on their visual qualities.

Quantitative Evaluation Results

10 de-noising methods have been tested, and their quantitative results are shown as below:

Methods	Natural Images PSNR	Natural Images PSNR-HVS-M	Synthetic Images PSNR	Synthetic Images PSNR-HVS-M
Wavelet Soft	26.675779	24.892400	18.323596	20.717732
Wavelet Hard	26.503098	25.253358	20.242995	22.673620
BM3D	29.465977	29.398808	26.115930	30.154313
Arithmetic Mean	24.622222	22.720622	15.153704	16.424611
Geometric Mean	21.755651	18.621081	10.419596	7.208516
Harmonic Mean	20.982258	17.587961	9.888079	6.510725
Positive ContraHarmonic Mean	21.125803	17.743720	11.782731	9.021239
Negative ContraHarmonic Mean	20.093378	16.481105	9.434632	5.951107
Median	24.893158	22.920620	14.881578	15.471420
Midpoint	16.780611	12.749748	12.161655	9.552054

Insights:

• BM3D and wavelet produce competitive results, which always rank the first and second respectively in each column of the results.
• Median filter generally outperforms the average filters, and Midpoint filter achieves the poor results.

8 wavelet filters have been tested using soft thresholding, and their quantitative results are shown as below:

Wavelet Filters (Soft)	Natural Images PSNR	Natural Images PSNR-HVS-M	Synthetic Images PSNR	Synthetic Images PSNR-HVS-M
Haar	23.622835	19.928527	17.451116	16.932536
Daubechies	24.647095	21.691431	16.872221	17.102421
Coiflets	24.219864	21.020080	16.626447	16.392394
Symlets	24.777326	21.940320	16.796563	17.104850
Fejer-Korovkin	24.014337	21.156380	17.347374	17.448833
Discrete Meyer	24.859222	21.967668	16.658263	16.888724
Biorthogonal	26.675779	24.892400	18.323596	20.717732
Reverse Biorthogonal	22.834260	19.172639	15.672878	14.445390

Insights: The Biorthogonal filter outperforms all other Wavelet filters in each column of measurement, which is why we choose Biorthogonal filter as the Wavelet filter when comparing Wavelet with other de-noising methods.

Qualitative Evaluation Results

We conduct the qualitative evaluation by examining the de-noised images of different methods, and get the following insights:

• BM3D and Wavelet are the best two denoising methods.
• BM3D can reduce much noise density, however it tends to smoothen the images and lose important details.
• Wavelet performs quite well in denoising natural images but cannot outperform BM3D from either quantitative or qualitative aspect.
• All the denoising methods get worse results for synthetic images than their results for natural images, and the rankings of different methods with respect to processing natural and synthetic images are different, which indicates the performance and strength of each method may also depend on specific image data.

Note: The de-noised images of different methods can be found in the “results” folder.