| Abstract [eng] |
In this research paper we have analyzed hyperspectral images, hyperspectral imaging technolo- gies and the noise origins in hyperspectral images. We have analyzed and implemented multiple HSI noise evaluation methods: correlation coefficient based (R1 and R2) and linear regression based (LMLSD and SSDC). To compare different noise evaluation methods we have implemented a simple median filter as described in section 1.3.1. For the analysis we have used 5 different noisy hyperspectral crop images captured from an airborne drone. We have concluded that in all noise evaluation methods the noise estimate parameter values were lower after filtering than in the original image, therefore, concluding that our designed noise evaluation methodology can be used to represent noise levels in HSI images. Furthermore, after analyzing the different noise evalua- tion methods we have concluded that the correlation coefficient based estimate methods and the linear regression based methods all give different results on how effective the filtering is. This was because the initial noise in the images was unknown, so we could not tell which of the noise evaluation methods provided the most accurate representation of the noise. However, we have still concluded that the linear regression based algorithm - LMLSD does not provide accurate re- sults in estimating the overall HSI data cube noise. The main reasons for the inaccurate results were identified as due to the noise not being fully homogeneous. However, the other implemented methods - SSDC, R1 correlation coefficient and pixel correlation R2 provided with similar results on determinig which are the most noisy bands and two of them were used to evaluate advanced filtering algorithms. More advanced filtering algorithms were implemented and compared. The best results were given by the BM3D algorithm which outperformed wavelet 2D, 3D and the basic median filter in all noise estimate parameters. More importantly it did not change the original signal LM values and the initial signal remained intact after filtering (full results in table 7). This was not demonstrated by any other method and this method is our recommendation for denoising HSI images. Direct 2D and 3D Wavelet Transformation based filtering was, also, implemented with different threshold functions and different wavelet forms. It was found that there are mini- mum dependencies on the selected wavelet form but big dependencies on thresholding functions. The best results were achieved with the λ threshold parameter and α threshold function (equations 1.23, 1.22). 3D DWT based filtering performed worse compared to 2D due to images not having homogeneous noise in all bands and, therefore, impacting the thresholding parameter accuracy. As expected, median filtered provided good SSDC LSD results but it had the biggest changes in signal LM values, changing the original useful signal information. Finally, Peak-Signal-To-Noise Ratio (PSNR) parameter (equation 1.12) was calculated to asess image quality for different filter- ing technologies. The results showed that the worst performing method was the wavelet 3D. Other methods provided simillar results with BM3D having the highest PSNR values in λ[550; 900] inverval and lowest in λ[400; 550] and λ[900; 1000]. This means biggest filtering was done the lowest and highest bands, and least filtering was done in mid-range. Therefore, concluding that the BM3D targeted the noisiest channels for filtering leaving the least noisy intact. |
