Enhancement of Degraded Image Based on Neural Network

Transcription

1 Engneerng scence Enhancement of Degraded Image Based on Neural Network Defa Hu School of Computer and Informaton Engneerng, Hunan Unversty of Commerce, Changsha , Hunan, Chna Zhuang Wu* Informaton College, Captal Unversty of Economcs and Busness Bejng , Chna *Correspondng author: Abstract Inevtably, mage degradaton s caused n the magng, reproducton, scannng, transmsson and dsplay. Ths paper conducts systematc analyss and research on the clearness processng of degraded mage based on BP neural network. The proposed scheme ncreases the output ranges of the hdden layer and the output layer, sets the varable step sze, accelerates the learnng speed of the neural network and avods the excessve correcton of the weghts and speeds up the convergence rate of the neural network. It can preserve more profound mage nformaton, greatly enhance the contrast of the degraded mage, effectvely strengthen the overall qualty of the degraded mage and obtan more deal mage clearness effect whle processng the degraded mage acqured under bad condtons. Key words: IMAGE CLEARNESS, BP NEURAL NETWORK, DEGRADED IMAGE 1. Introducton The mage we record usually has certan degree of degradaton, ncludng some pxel dsplacement, object dstorton and dstance rato mbalance n the mage due to the mperfect actual magng system, the mpact of transmsson meda, the relatve moton between the scenery and the magng system and the random envronment noses n the formaton, transmsson and recordng of the mage[1]. In fact, ths s the so-called mage degradaton. The purpose of mage clearness restoraton processng s to process the degraded mage to restore t nto the orgnal mage and ths s the foundaton of mage processng, pattern recognton and machne vson[2]. The tradtonal mage clearness restoraton s faced wth the computaton of hgh-dmensonal equatons, whch has a heavy computaton load, requres the assumpton to satsfy the generalzed statonary process n the restoraton or lacks complete theoretcal foundaton and unfed desgn method[3]. Ths s the fundamental reason why the mage restoraton problem wth extensve applcaton value can t be solved satsfactorly. Neural network has certan advantages n the Metallurgcal and Mnng Industry, 2015, No

2 restoraton of mage clearness because the mage clearness based on neural network doesn t need to assume that the mage meets the wde-sense statonary process and that t s easy to realse[4]. By analyzng the features of degraded mage and havng a great number of tranngs, neural network method can accurately dentfy and extract the fuzzy regons n the locally-moved fuzzy mage, search the fuzzy features n the mage, set the parameters, process the fuzzy regons and get the restored mage n the complcated backgrounds and unclear gray features[5]. Frstly, ths paper desgns the hdden layer, the output layer and the step sze of BP neural network n accordance wth the clearness restoraton of degraded mage. Secondly, t gves the basc procedures of the algorthm n ths paper and the steps of the restoraton of mage clearness. Fnally, t s the experment smulaton and analyss. 2. BPAlgorthm BP neural network algorthm s a backpropagaton algorthm, the operatng base of whch s mult-layer feed-forward neural network. Ever snce t was proposed n the 1980s, t has been attractng ncreasng attenton, so far, t has been wdely appled n varous forefront felds Introducton of BP Algorthm As one of the layered network models, BP neural network model ncludes nput layer, output layer and multple hdden layers. Besdes, there are several unts n every layer of BP neural network model and the connecton between the unts s realzed by drected weghted edge. Attenton shall be pad to several problems here: 1) The network s feed-forward and sngle, that s to say, every feedback can only be sent to the front neghborhood output layer or hdden layer but not to span and propagate backwards. 2) The network s fully connected. In other words, t s n a one-to-many relatonshp, namely that the unts n every layer are connected wth all unts n the prevous layer through drected weghted edge. Therefore, as long as there are enough hdden layers n the mddle, the lnear threshold functon n the mult-layer feed-forward neural network can approach any functon suffcently. Besdes, before the neural network tranng starts, the structure of the neural network must be confrmed, namely the followng shall be confrmed: the unts n the nput layer, the number of hdden layers, the unts of every hdden layer and the unts of the output layer. However, there sn t a specfc theoretcal bass for us to follow as for how to confrm the number of nodes n every hdden layer and the number of network layers [6] Process of BP Algorthm In the above dscusson, t has been confrmed BP neural network model s one of the mult-layer feed-forward neural network models.its structural pattern must be confrmed before utlzng BP neural network model (t s certan that such structure may change after that. Then dvde the orgnal data provded, ncludng tranng data, test data and nspecton data. In the calculaton, obtan the calculaton error between the data and that of the node n the forward layer and adjust ts weght and threshold so that the nput data and output data satsfy certan mappng relatonshp gven before[7]. It can be seen that by prncple, BP neural network model can be dvded nto three parts. Forward propagaton nput: Calculate the net nput of the unt I = wo +θ and the data to be used s the bas j j j I between the lnear combnaton and unt. Use actvaton functon actvaton functon n the net nput of the unt and get the unt output 1. O = j 1 + e I Calculate the error Error calculaton reflects the network predcton error through the weght updates and the bas and back propagate the predcted error. Calculate the error E rrj of the unt j n the output layer wth the followng formula: Errj = Oj (1 Oj )( Tj Oj ) (1) Update the weght and bas The update methods nclude nstance update and perodc update. The former update of weght and bas are made after processng a data. In ths way, t has excellent tmelness but ts parallel processng capacty s not so strong and that s why perodc update comes nto beng. The basc dea of perodc update s to update the weght and bas after processng traned and centralzed samples. There are three termnaton condtons: the frst s that the varaton of all w j n the prevous perod s smaller than a certan gven threshold, the second s that the percentage of the samples whch are not correctly classfed n the prevous perod s smaller than a certan threshold and the thrd s that t has exceeded the pre-assgned number of perods[8]. 2.3 Standard Formula of BP Algorthm It s nevtable to dscuss the nput and output of the nodes n the BP neural network 282 Metallurgcal and Mnng Industry, 2015, No. 4

3 model.there are no specfc conclusons to follow n the exstng academc materals n the selecton of the ntal neural network parameters such as the weght w, the devaton value θ and the learnng rate of that network model[9]. The nput and output calculaton formula (1) The value of the nput node n the nput layer: x j (2) The output of the hdden node n the hdden layer: y = f( wx θ ) j j (3) The output of the output node n the output layer: o = f( T θ ) node: t j The modfed formula of the output layer (1) The expected output of the output (2) Error control: p k ξr k = 1 E = e < (3) Error calculaton: Errj = oj (1 oj )( Tj o j ) (4) Weght modfcaton: Tj ( k+ 1) = Tj ( k) + lerrjo (5) Threshold modfcaton: θj ( k + 1) = θj ( k) + lerrjo The modfed formula of the hdden layer (1) Error calculaton: E = o (1 o ) E w rrj j j rrk kj k (2) Weght modfcaton: wj ( k+ 1) = wj ( k) + lerrjo (3) Threshold modfcaton: θ ( k + 1) = θ ( k) + le o j j rrj 3. Establshment of BP Neural Network 3.1 Desgn of Transfer Functon Transfer functon s an mportant part of BP network and we usually use S-type logarthmc or tangent functon. Before the nput, make a quantzed unfcaton on the orgnal data to make the nput vectors wthn the range of [- 1,1] and meet the value requrements of the above transfer functons. Therefore, ths paper selects tan sg and logsg as the transfer functons of the hdden layer and the output layer respectvely [10]. tan sg s a hyperbolc tangent S-type transfer functon wth ts graph as ndcated by Fg.1. Fgure 1. Hyperbolc tangent s-type transfer functon logsg s the S-type logarthmc functon and ts call format s: A = log sg( N ) nfo = log sg (code) In here, N s S-dmensonal nput vectors and A s the functon return value wthn the range of (0,1). nfo = log sg (code) returns dfferent nformaton accordng to the dfferences of code value. See ts functon graph as Fg.2. Fgure 2. S-type logarthmc functon 3.2.Desgn of Intal Weght Intal weght s of great sgnfcant n the neural network because t determnes the ntal state of the error. Normally, the smaller the desgn of the ntal weght, the better snce n ths way every neuron n the neural network can be close to unform dstrbuton n that layer. Ths paperhas few requrements on the parallel capacty and error analyss of BP neural network, however, snce the predcton result s based on short-term trend, t has hgher accuracy requrements. Ths paper uses Functon nt provded by Matlab. An ntal value randomly generates Metallurgcal and Mnng Industry, 2015, No

4 among [-1,1]. It uses the characterstcs of BP neural network n the entre process, adjusts ts numercal value by calculatng the error and fnalzes the numercal value of ntal weght[11]. 3.3 Desgn of Tranng Functon The data collected n ths paper are all wthn the range of [0,1] after normalzaton processng and the dfferences between every numercal values are small, therefore, the tranng functon ths paper selects s trandx. In accordance wth the tme, number and accuracy requrements on the tranng data, ths paper defnes the tmes of tranng as and sets the tranng target error and learnng rate as and 0.01 respectvely. See the detals as followsn Fg.3: Weght correcton Input layer Hdden layer Output layer Fgure 3. Tranng process of BP network Error Tranng 3.4 Desgn of Performance Functon In ts research, n order to make the model unversal and applcable, t mproves ts generalzaton ablty on new samples to reduce tranng error. The feed-forward network error performance functons used n ths paper are Mean Square Error mse: N N F = mse = e = ( t a) (2) N = 1 N = a Although there s another performance functon, whch s the modfyng network error performance functon msereg, namely: msereg = r * mse + (1 r) msw (3) In here, r s the error performance 1 2 adjustment rato and = n msw x j. n j= 1 The above has form the complete desgn and study process of a complete BP neural network model[12]. 4. The Flow and Steps of BP Neural Network Algorthm The mage clearness based on BP neural network algorthm ncludes the followng steps: (1) Read the mage, obtan ts pxel matrx, extract the gray value of every pxel pont, computes the membershp of pxel ( mn, ) n the matrx by usng the membershp functon and set the nput vector P and the target vector. (2) Intalze BP neural network and set the learnng rate η (0) of the frst sample tranng. The tranng lasts from the nput of the samples tll the network error reaches the set value or the maxmum number of tranngs has reached and preserve the weght and threshold. (3) Read n the mage to be processed. The mage to be adopted here s the football mage wth a gray scale of 256. Degrade the orgnal mage through 5 5 functon and ts form s: H = / 32 (4) (4) Defne the range of the gray value of the degraded mage Y wthn [ 0.5,0.5],.e. the value range of every neuron x s wthn [-0.5, 0.5], so as to reduce the error whch s ntroduced when forcng x of every teraton nto the value Y range to certan extent: Y = 0.5, the ntal 255 T value x(0) = H Y, t = 0 and the admssble error 5 of the network convergence s err = 10. (5) Select the mage to be sharpened, tran the nput vector wth the well-traned BP neural network. And the fnal output vector s the restoraton processng result of the mage. (6) Restore the mage restoraton result nto the mage grayscale matrx n the form of vector and dsplay the result. The key flow othe mproved method s ndcated n Fg.4: 284 Metallurgcal and Mnng Industry, 2015, No. 4

5 Start Buld BP neural network, nput tranng data and select the membershp functon BP neural network computaton Compute the error Reach the target or not? No Yes Read n the orgnal mage Tranng s over, preserve the network and output the basc parameters Perform mage degradaton processng and acqure degradaton matrx Extract the mage feature vector and compute the weght and threshold by usng the parameters obtaned from BP neural network tranng. BP neural network trans the vectors acqured n the mage. Compute the result of the mage restoraton vector and restore t nto mage gray matrx n the form of vector Output the preserved restored mage. End Fgure 4. The man flow of mproved method The selecton of the learnng rate η s very mportant. If η s too bg, the weghted coeffcent may not be converged because of repeated vbraton, f η s too small, the learnng rate can be relatvely slow, causng too much tme n network convergence, therefore, the nerta coeffcent αs usually ntroduced to accelerate the network convergence when η s small and to assst n the network convergence when η s bg[13,14]. Another problem of BP network s that the system may be trapped nto certan local mnmum, or certan quescent pont or vbraton among these ponts n the learnng process. Under these crcumstances, the system wll have huge errors no matter how many teratons have been performed. Therefore, n the learnng process, avod the system to be trapped n certan local mnmum ponts and the ntroducton of nerta tem may avod the network to be trapped n vertan local mnmum[15]. 5. Experment Smulaton and Analyss The orgnal clear mage adopted n the smulaton experment s the 128x128 football mage wth a gray scale of 256. Fg.5(C) s the restoraton result of the least square method and Fg.5 (D) s the restoraton result of BP network. It can be seen from the comparson of the above experment results that BP neural network has greatly mproved the qualty of the restored mage wth well-preserved mage detals, comfortable sense of feelng to human eyes n the processed mage and obvous effect. The method ntegratng tan sg functon and log sg functon has provded the network wth stronger functon approxmaton capablty and the obvous mprovements n the mage restoraton has proved Metallurgcal and Mnng Industry, 2015, No

6 that the method n ths paper s better than the tradtonal least square method. (a) Standard mage (b) Degraded mage (c) Least square method (d) BP neural network Fgure 5. Comparson of restoraton of mage clearness 6. Concluson In varous mage systems, the mage s usually degraded because of mage transmsson and converson such as magng, reproducton, scannng, transmsson and dsplay. In order to mprove the mage qualty, ths paper has nvestgated the clearness of degraded mage based on BP neural network from the aspects lke the transmsson functon, the network learnng rate and the tranng algorthm of BP neural network. And the experment smulaton n the fnal part has verfed the effectveness of the method of ths paper. Acknowledgements Ths work was supported by the Bejng Phlosophcal Socal Scence Project (No.14SHB015), the Bejng Muncpal Educaton Commsson Foundaton of Chna (No. SM ) and the Natonal Natural Scence Foundaton of Chna (Grant No: ). References 1. Olver Jovanovsk (2014) Convergence Bound n Total Varaton for an Image Restoraton Model. Statstcs & Probablty Letters, 90(7), p.p Zhuang Wen Wu, Langrong Zhu (2013) Car Informaton Bus Image Restoraton usng Multwavelet Transform Algorthm. TELKOMNIKA Indonesan Journal of Electrcal Engneerng, 11(10), p.p Donghong Zhao (2014) Total Varaton Dfferental Equaton wth Wavelet Transform for Image Restoraton. TELKOMNIKA Indonesan Journal of Electrcal Engneerng, 12(6), p.p Pablo Ruz, Hram Madero-Orozco, Javer Mateos, et al. (2014) Combnng Posson Sngular Integral and Total Varaton Pror Models n Image Restoraton. Sgnal Processng, 103(10), p.p Jahedsaravan, M.H. Marhaban, M. Massnae (2014) Predcton of the Metallurgcal Performances of a Batch Flotaton System by Image Analyss and Neural Networks. Mnerals Engneerng, 69(12), p.p M. Sartha, K. Paul Joseph, Abraham T. Mathew (2013) Classfcaton of MRI Bran Images usng Combned Wavelet Entropy Based Spder Web Plots and Probablstc Neural Network. Pattern Recognton Letters, 34(16), p.p Metallurgcal and Mnng Industry, 2015, No. 4

7 7. M. Monca Subashn, Sarat Kumar Sahoo (2014) Pulse Coupled Neural Networks and Its Applcatons. Expert Systems wth Applcatons, 41(8), p.p Seung-Ho Kang, Jung-Hee Cho, Sang-Hee Lee (2014) Identfcaton of Butterfly Based on Ther Shapes when Vewed from Dfferent Angles usng An Artfcal Neural Network. Journal of Asa-Pacfc Entomology, 17(2), p.p D. Jude Hemanth, C.Kez Selva Vjla, A.Immanuel Selvakumar, et al. (2014) Performance Improved Iteraton-free Artfcal Neural Networks for Abnormal Magnetc Resonance Bran Image Classfcaton. Neurocomputng, 130(23), p.p Roberto Vega, Gldardo Sanchez-Ante, Lus E. Falcon-Morales,et al. (2015) Retnal Vessel Extracton usng Lattce Neural Networks wth Dendrtc Processng. Computers n Bology and Medcne, 58(1), p.p Sddhartha Bhattacharyya, Pankaj Pal, Sandp Bhowmck (2014) Bnary Image Denosng usng A Quantum Multlayer Self Organzng Neural Network. Appled Soft Computng, 24(11), p.p Anna Aprle, Govanna Castellano, Gacomo Eramo (2014) Combnng Image Analyss and Modular Neural Networks for Classfcaton of Mneral Inclusons and Pores n Archaeologcal Potsherds. Journal of Archaeologcal Scence, 50(10), p.p A.Bouhamd, R. Enkhbat, K. Jblou (2014) Condtonal Gradent Tkhonov Method for a Convex Optmzaton Problem n Image Restoraton. Journal of Computatonal and Appled Mathematcs, 255(1), p.p Rachd Hedjam, Mohamed Cheret (2013) Hstorcal Document Image Restoraton usng Multspectral Imagng System. Pattern Recognton, 46(8), p.p Ratnakar Dash, Banshdhar Majh (2014) Moton Blur Parameters Estmaton for Image Restoraton. Optk-Internatonal Journal for Lght and Electron Optcs, 125(5), p.p Metallurgcal and Mnng Industry, 2015, No