The Research On Filtering Technology Of Stripe Noise On Hyperspectral Image

With the continuous development of science and technology, cognitive competence of people has been improved. One of the most iconic achievements of science and technology in the 20th century is the rise of remote sensing technology. Especially in the early of 18th century, with the continuous development of spectral imaging technology, optical remote sensing has entered a new phase-the phase of Hyperspectral Remote Sensing. Due to high spectral resolution, the high spectral remote sensing images have been widely used in the field of civil and military. Hyperspectral image can provide a wealth of spatial information with the characteristic of "combine spectrum into picture", multi-band and large amount of data. However, stripe noise the prevalently exists in Hyperspectral image. There is a considerable difference between stripe noise and random noise. The stripe noise is zonal distribution, and has a certain width.According to the characteristics of hyperspectral image and noise statistics law, domestic and foreign scholars proposed many stripe noise filtering methods,including histogram matching method, moment matching metnod, and some other methods on transform domain about Fourier Transform and Wavelet Transform. Until now, parts of the existing denoising methods have been improved by scholars. However, Most of these methods are stick to the characteristics of imaging spectrometer, the width and distribution of strip noise. They do not have general applicability and moderate complexity of the algorithm. Therefore, we need to do further study on hyperspectral image to explore a suitable band noise filtering method on hyperspectral image. This method can reduce the request of surface features, and furthest maintain the texture details with filtering out the stripe noise. In this paper, the mechanism of band noise and filtering algorithms on Hyperspectral image was in-depth study. The contents were as follows:1. An an improved neighbor interpolation algorithm which is based on moment matching metnod was proposed. By analysising on discriminant function, a new discriminant formula has been constructed. To some extent, new method makes up for the disadvantage of "information on average" of the original method. According to neighborhood different information elements having different impact on current information element, we define different correlation coefficients for neighborhood different information elements. Simulation results show the improved interpolation algorithm in the neighborhood is better than the original method of in evaluating indicator. 2. Because of different imaging mechanism, different widths stripe noise may exist in hyperspectral image simultaneously. This paper presents a method combining moment method with interpolation menthod, which is fit for the presence of any width of band noise. The method can automatically filter strip noise with any width on AVIRIS image, overcoming the disadvantage of original method. Simulations on AVIRIS image show that the proposed method can not only effectively filter the band noise of the image, but also be able to ensure image quality.