Advanced cience and Technology Letter Vol.11 (AT 16) pp.85-89 http://dx.doi.org/1.1457/atl.16. Iage Denoiing Baed on Non-Local Low-Rank Dictionary Learning Zhang Bo 1 1 Electronic and Inforation Engineering Changha Noral Univerity Changha 411 China Abtract. For hyperpectral reote ening iage denoiing thi paper propoed iage denoiing baed on non-local low-rank dictionary learning. The baic idea of algorith i to ue trong relativity of all wave band of hyperpectral reote ening iage with local elf-iilarity and local parity of iage to iprove the denoiing perforance. Firt of all cobined with the trong relativity non-local elf-iilarity and local parity non-local low-rank dictionary learning i etablihed. Then iterative ethod i ued to olve the odel to get redundant dictionary and parity to repreent coefficient. Finally redundant dictionary and parity i ued to expre retored iage of coefficient. Copared with the exiting advanced algorith by aking full ue of trong relativity each band of hyperpectral iage it ake the algorith obtain the inforation on detail to well keep the hyperpectral reote ening iage to iprove the viual effect. Experiental reult verify the effectivene of the algorith in thi paper. Keyword: Iage denoiing; Hyperpectral; Reote ening iage; Low-rank; Dictionary learning; pare repreentation 1 Introduction Conidering the ue of trong relativity between iage of each band of hyperpectral reote ening and cobined with the nonlocal elf-iilarity and local parity of iage thi paper propoed iage denoiing baed on non-local low-rank dictionary learning. The algorith firtly etablihe a non-local low-rank dictionary learning odel and then contruct the iage that i correponding tp denoiing algorith to olve the odel. Due to introducing the intrinic characteritic of trong relativity of band iage it ake thi algorith effectively keep texture and detail of each band iage while denoiing. IN: 87-133 ATL Copyright 16 ERC
Advanced cience and Technology Letter Vol.11 (AT 16) Dictionary Learning The pace width of pectral reote ening iage i et a W the height a H band dienion a. Conidering the iage noie i additive noie the obervation odel i Y XΕ (1) W H Wherein Y R i the iage polluted by the noie W H X R i the W H original iage Ε R i additive noie. n K The dictionary learning technology. Given a et of aple β=( β1 β K ) R n the purpoe of the dictionary learning i finding a redundant dictionary D R to ake each aple able to be expreed a the pare repreentation atrix by parity Γ=( Γ Γ Γ K ) R which i expreed a 1 K β DΓ Dx F t.. Γk T k () The coon evaluation ethod for dictionary learning proble () include MOD algorith K-VD algorith[15] or online dictionary learning algorith[16] etc. 3 Iage Denoiing Algorith Baed on Non-Local Low-Rank Dictionary Learning Thi algorith in thi paper firtly etablihe a non-local low-rank dictionary learning odel then contruct the correponding algorith for olving the odel to get a redundant dictionary and pare repreentation coefficient and finally it ake ue of the reult to retore iage. Data tructure of hyperpectral reote ening iage i a cube. Aug that the iage contain band iage to divide hyperpectral iage into N overlap fullwave band cube data with the ize of n n (hereinafter referred to a the cube) repreented a pn ( n1 N) and dictionary learning i expreed a: D αn.. t N n1 1 α p n n n T Dα (3) 86 Copyright 16 ERC
Advanced cience and Technology Letter Vol.11 (AT 16) Wherein p R n n i the expreion for of the iage block vector of the brand in the n cube D i the redundant dictionary α n i pare repreentation coefficient. Conidering the ue of nonlocal elf-iilarity k ean clutering algorith i ued to ake N cube data into K clae and then take advantage of cube data of each cla to learn to get ub-dictionary and ue ub-dictionary to expre data in the cla. Cube data in each cla are repreented a M k and the nonlocal dictionary learning odel in the k cla i repreented a α.. t M k 1 1 α p ( k) ( k) k T D α (4) cla. Wherein p i the iage block vector of brand of the cube in the k i the ub- dictionary in the k cla repreentation coefficient. ince α i the correponding pare p can be expreed by ub-dictionary D k o α T can be et. Therefore the nonlocal dictionary learning odel in the k cla (4) i equivalent to α 1 1 ( k) ( k) k p D α (5) Becaue all brand iage of hyperpectral iage have trong correlation the coefficient atrice correponded by the full-wave data of the n cube data ( k) ( k) ( k) α α1 α i a low-rank atrix. The low-rank contraint of the atrix α are added to the nonlocal dictionary learning odel () to get the nonlocal low-rank dictionary learning odel a α p D α α (6) k rank 1 1 1 Wherein i the weighted paraeter rank i atrix rank. Matrix rank. Matrix rank often ue nuclear nor (the u of atrix eigenvalue) to approxiate the nonlocal low-rank dictionary odel i equivalent to Copyright 16 ERC 87
Advanced cience and Technology Letter Vol.11 (AT 16) α p D α α (7) k 1 1 1 The odel i iplified to get α 1 p α α (8) F Wherein the atrix ( k) ( k) ( k) ( ) p p1 p α α1 α naely p k n i the full-wave data of the cube i the k cla αn i the correponding coefficient atrix 4 Detailed tep and Analyi of the Algorith Thi ection will decribe the detailed tep of the algorith in thi paper a hown in algorith 1. Aalgorith 1: Denoiing baed on non-local low-rank dictionary learning W H Algorith input: Y R containing noie hyperpectral iage Initialization: initial dictionary D () ; tep 1 aking the hyperpectral iage divided into overlapping cube all-band data tep aking all p n divided into K clae by the k-ean clutering algorith; tep 3 by olving the nonlocal low-rank dictionary learning odel (9) in the K cla the coefficient atrix of p n αn and the correponding dictionary D n are gained; tep 4 etiating each cube all-band data pˆn D n α n getting the average of joining the overlapping part according to location retoring the iage ˆX Output reult: de-noied iage ˆX. Note of Algorith 1: () () (a) The election of initial dictionary D : for the initial dictionary D it uually chooe quickly ipleented data-type dictionary uch a DCT dictionary all wave dictionary. (b) The election of paraeter : the nonlocal low-rank dictionary learning proble are regarded a a ulti-objective optiization proble which can be obtained through the ethod -. 88 Copyright 16 ERC
Advanced cience and Technology Letter Vol.11 (AT 16) 5 Concluion For hyperpectral reote ening iage denoiing thi paper propoed in denoiing baed on the nonlocal low-rank dictionary learning. The core idea of the algorith i to ake ue of trong relativity of the hyperpectral iage of variou hand in the proce of denoiing cobined with local parity and non-local elf-iilarity of iage. Due to the ue of trong relativity of the hyperpectral iage of variou hand thi algorith can ake denoiing achieve better reult. The experiental reult how that the PNR value of the algorith to retore iage i higher than the exiting advanced algorith and can keep the detail inforation of the iage well o a to iprove the viual effect. Reference 1. Zhang X. Han Y. Hao D. Lv Z.: ARPP:Augented Reality Pipeline Propect yte. th International Conference on Neural Inforation Proceing (ICONIP 15) Itanbul Turkey. In pre.. Hu J. Gao. Z.: Ditinction iune gene of hepatiti-induced heptatocellular carcinoa [J]. Bioinforatic (1) 8(4): 3191-3194. 3. Wang K.: Overcog Hadoop caling Liitation through Ditributed Tak Execution 4. Zhang. Zhang X. Ou X.: After we knew it: epirical tudy and odeling of coteffectivene of exploiting prevalent known vulnerabilitie acro IAA cloud. Proceeding of the 9th ACM ypoiu on Inforation coputer and counication ecurity. ACM (14) Copyright 16 ERC 89