Você está na página 1de 2

Sonu Sainju 11130264 1)

Assignment #2 Date- 2/6/2012

a) An image is chosen and different noises are added to it. Table 1 below shows the PSNR of the noisy images. Noise Model Gaussian Salt & Pepper Poisson Exponential
Table 1

PSNR 21.88 16.69 31.08 44.95

All of the noises were implemented using the in built functions of MATLAB except exponential noise. To model noising of an image by exponential noise, simple addition was performed as: I+N Where, I is the matrix of image pixels and N is the matrix whose elements are exponentially distributed. On viewing the image, salt and pepper and Gaussian noise seem to degrade the image more than other noises and the similar results can be deduced from the PSNR of the noisy images. b) PSNR of the denoised images obtained by applying denoising filters (3x3 kernel size and zero padding): Noise Gaussian Salt & Pepper Poisson Exponential Uniform Rectangular 28.05 25.31 36.38 38.03
Table 2

PSNR Uniform Circular 27.14 23.47 36.84 41.10

Median 28.88 39.25 36.09 39.83

c) PSNR of the denoised images obtained by applying denoising filters (5x5 kernel size and zero padding): Noise Gaussian Salt & Pepper Poisson Exponential Uniform Rectangular 28.43 27.47 33.34 33.59
Table 3

PSNR Uniform Circular 28.61 26.77 35.47 36.17

Median 31.26 36.10 35.26 36.40

d) When viewing the denoised images in case of median filtering, there werent much differences between denoised images implemented using 3x3 and 5x5 kernel size filters, except gaussian noise added image. The image quality or removal of Gaussian noise is more with filter using 5x5 kernel size than 3x3 kernel size. Same information can be deduced from comparing PSNR s from above two tables. However, in case of poisson and exponential noised images, using 5x5 kernel sized filters degraded PSNR of images than smaller sized filters. e) On viewing the images of the median filter, it seems that it has best results in case of salt & pepper noised image. It was also noted that increasing the size of the kernel in these filters increased the PSNR and also the quality of the image in the particular case of salt & pepper noise. This kind of filter works well with other kind of noised images as well. In case of the averaging filters, the removal of the salt & pepper noise was fairly better with large kernel size but it was seen that this resulted more in the blurring of the image and thus the picture quality degraded. But, with Gaussian noised images, this kind of filters had better picture restoration results. When the kernel size was increase in this case also, there was more removal of the noise but again the compromise had to be made with thee image quality due to blurring. f) PSNR of the denoised images obtained by applying denoising filters (3x3 kernel size and replicated boundary): Noise Gaussian Salt & Pepper Poisson Exponential Uniform Rectangular 28.06 25.26 36.33 38.03
Table 4

PSNR Uniform Circular 27.15 23.44 36.764 41.10

Median(symmetric) 28.83 39.34 36.05 39.83

As expected, the PSNRs of this case were very similar to those of table 2. The quality of the images was also seen to be quite same. However, one major difference observed was that, on the zeropadded case, images had dark boundaries which were completely removed in the case of replicated boundaries.

Você também pode gostar