Research On Seismic Denoising Methods Based On Block Similarity

Random noise suppression is one of the most basic flows in seismic data processing.A good denoising result is also the basis of subsequent processing and interpretation.At present,most of the commonly used methods for suppressing random noise in seismic data denoise the whole section at once,and are lack of the use of local block features.However,the denoising methods based on block similarity in image processing fully solve this problem.Two image denoising methods based on block similarity,non-local mean(NLM)and block matching and three-dimensional collaborative filtering(BM3D),are introduced in this paper.On the basis of the detailed description of the two algorithms,the influence of NLM algorithm's search area size,Gauss kernel variance and filtering parameter h,as well as the hard threshold value and preset variance of BM3D algorithm on the final denoising results is discussed.To solve the problem of low speed and time-consuming calculation of NLM,slope-constrained NLM algorithm and adaptive frequency NLM algorithm are proposed in this paper.The former mainly relies on the prior local slope information,and adaptively separates an inclined zone in the square area that needs to be searched all the time.It reduces the number of points that need to be calculated.At the same time,according to the slope information,an adaptive kernel function is designed to make it more consistent with the fact that the events of seismic data are mostly horizontal textures,so as to optimize the weight distribution.On the premise of achieving the same or slightly better denoising results,this method only takes half of the time of the traditional method.The latter introduces the original algorithm used in space-time domain into frequency-space domain(f-x domain),which converts the two-dimensional problem into multiple one-dimensional problems and uses NLM to denoise each frequency slice,and uses conjugate symmetry of Fourier transform and band-limited characteristics of signal frequency domain to reduce the computational complexity.In addition,the algorithm sets different filtering parameters according to the energy of different frequency slices,which can protect the signal frequency band with weak energy.The algorithm takes into account both the advantages of frequency domain processing and non-local characteristics,and achieves a computational speed of tens to hundreds of times that of traditional NLM.Aiming at the problem that BM3D filtering tends to damage complex structural signals,this paper combines BM3D with local similarity threshold,which not only makes use of the suppression effect of BM3D which is far superior to traditional methods in initial denoising,but also can accurately use local similarity to recover the signal part from the noise profile and reject the noise part.Compared with the traditional methods,the algorithm has made some progress in signal-to-noise ratio and fidelity.For these different improved methods,different synthetic and real data are used to test their effectiveness.The results show that both the calculation speed and the quality of the final section have been improved to a certain extent.