Mcrowave Dversty Imagng Compresson Usng Bonspred Neural Networks Youwe Yuan 1, Yong L 1, Wele Xu 1, Janghong Yu * 1 School of Computer Scence and Technology, Hangzhou Danz Unversty, Hangzhou, Zhejang, P.R Chna, 310018 * Department of mechancal Engneerng, Hunan Unversty of Technology, Zhuzhou, Hunan, P.R Chna, 412008 y.yw@163.com Abstract. In ths paper a novel bo-nspred neural network (BINN) has been created for mcrowave dversty magng compresson. Ths work focuses on, frstly ncorporatng BINN and back propagaton learnng nto a multresoluton, and secondly, tranng the hghpass channel for perfect reconstructon so that the BINN flterbank wthn realstc codng condtons. Experments have shown the robustness and accuracy of the BINN mplementaton of mcrowave magng compresson. Keywords: Bo-nspred, Neural Networks, Mult-resoluton, Image Compresson 1 Introducton Mcrowaves are used n the remote magng radar systems to obtan the nformaton about objects of nterests n many applcatons. The goal of mcrowave dversty magng compresson s to reduce rrelevance and redundancy of the mage data n order to be able to transmt or store data n an effcent form. The mportant propertes of mage compresson schemes ncludes scalablty, regon of nterest codng, meta nformaton and Processng power[1-2].the mcrowaves magng mechansm can be analyzed usng electromagnetc wave scatterng and lnear system theory. The rest of the paper s organzed as follows: Secton II gves a bref revew of the Bo-nspred neural networks(binn) approach for mcrowave dversty magng compresson. Secton III analyzes experments and dscussons.fnally presents the conclusons and future work was gven n secton IV. NGCIT 2013, ASTL Vol. 27, pp. 20-24, 2013 SERSC 2013 20
Proceedngs, The 2nd Internatonal Conference on Next Generaton Computer and Informaton Technology 2 Bo-nspred Neural Networks Approach for Mcrowave Dversty Imagng Compresson 2.1 Bo-nspred Neural Networks The structure of the proposed bo-nspred neural network s shown n fg.1.whch has been created by ourselves based on the presented works of others [3]. The nput layer can be expressed as a convoluton: dx dt = b, θ = 1, ρ = 0.2sn 2πt (1) where x, θ represent nternal potental and the threshold respectvely, and ρ s the relaxaton level of neural networks I, and b are the bnary nput. For the output layer can be formulated as[4] : y ( t) = y ( t d (, j)) (2) b j = 1b = 1 l n= n0 y ( t) = δ ( t n) (3) where d l (,j) s the lateral dstance and n 0 s postve nteger. The expresson for the output of the fnal summng node can be defned as: = = = y( t) = δ ( g dl (, j) n (4) n n0 b j 1 b 1 Fg. 1. The proposed bo-nspred neural network structure 21
Mcrowave Dversty Imagng Compresson Usng Bo-nspred Neural Networks 2.2 The Structure of Mcrowave Dversty Imagng Compresson Usng BINN The structure of mcrowave dversty magng compresson usng bo-nspred neural networks s shown n fg.2. Ths mcrowave dversty magng compresson s derved from some predcton on object movements n vertcal and horzontal drecton[5].the fnte frequency band [p 1,p 2 ] can be measured through a band-pass flter wth the transfer functon. Fg. 2. The structure of Mcrowave Dversty Imagng compresson Usng BINN In block-based moton estmaton (ME), we dvde an mage nto a number of blocks of pxels wth the assumpton that each block has a sngle moton. In the dversty spatal doman whch s correspondng to the tme doman, a fnte frequency band [p 1,p 2 ] for a gven angle can be defned as : or () = hr () or () (5) Where or () represents the deal range-profle; or () s the reconstructed rangeprofle; and hr () s the spatal mpulse whch relates to the system expressed: jrp h() r = H ( p) e dp (6) 1 2π + We use lnear block transform codng where the mage can be subdvded nto blocks and flter bank s terated to create a pyramd type representaton of the mage at dfferent resolutons. 22
Proceedngs, The 2nd Internatonal Conference on Next Generaton Computer and Informaton Technology 3 Experments and Dscussons The test condtons are nclude: all vdeo sequences under the wndows XP operatng system and mage s YUV=4:2:0, Input Fle= vdeo sequence needed to test, FramesToBeEncoded=100, FrameRate=30.0, SourceWdth=176, Sourceheght=144, UseHadamard=1, SearchRange=16, UseFME=1, NumberReferenceFrames=5, QP=28, and other parameters are default settng. The grd of neurons wth 40 rows by 200 columns ((40 200) s chosen for the specfc mage segmentaton mplementaton. After the process of adaptaton of the bo-nspred neural network, the weghtng vectors of the nput neurons wll have values dentcal to the approprate ponts. The neghbourng neurons are confned to a wndow of 3 3 neurons throughout the network tranng[6]. The test consequence s shown n Table.1. where the proposed method reduces both encodng tme and compresson tme whle mantan relatvely stable PSNR and bt rate. As shown n fg.3, our optmzed algorthm can reduce both the total encodng tme and moton estmaton tme as compared to UMHexagonS. Table 1. Test consequences UMHexagonS Our optmzed algorthm Vdeo sequence PSNR bt rate Enc.T ME.T PSNR bt rate Enc.T ME.T (db) (kbt/s) (s) (s) (db) (kbt/s) (s) (s) Moble 34.12 263.62 358.412 159.033 33.10 266.45 344.361 150.382 Coastguard 34.38 169.42 290.847 155.676 34.28 169.18 289.468 150.267 Slent 36.23 66.73 268.368 125.548 36.11 66.64 246.085 113.512 Contaner 36.43 32.15 234.477 102.105 36.43 31.13 230.768 100.936 Hghway 37.65 53.28 222.383 110.121 37.65 53.58 218.271 108.393 Foreman 37.65 53.28 219.588 108.371 37.61 54.04 214.702 105.316 (a) Slent_qcf.yuv Sequence (b) Contaner_qcf.yuv Sequence Fg. 3. Varous Test Sequence 23
Mcrowave Dversty Imagng Compresson Usng Bo-nspred Neural Networks 4 Summary of the Research and Future Work A bo-nspred neural networks (BINN) flter bank consstng of the analyss low and hghpass has been desgned. The tranng algorthm that consders the analysssynthess system has been developed and successfully mplemented. Experments have shown the robustness and accuracy of the BINN mplementaton of mcrowave dversty magng compresson. It s hoped that ths research wll open up more possbltes applcaton for future researcher n bo-nspred neural network and mage codng compresson. Acknowledgment. The work was supported by the Natonal Hgh Technology Research and Development Program of Chna (863 Program, No.2011AA01A107) and NSFC (Granted No. 61272032). References 1. Luthra, A., Sullvan, G. J., & Wegand, T. Introducton to the specal ssue on the H.264/AVC vdeo codng standard. IEEE Transactons on Crcuts and Systems for Vdeo Technology, Vol. 7-13 (2003), p.557-559. 2. Sullvan, G. J., & Wegand, T. Vdeo compresson from concepts to the H.264/AVC standard. Proceedng of the IEEE, Vol. 1 (2005), p. 18-31. 3. Maas W. and Sontag E.D. Neural systems as nonlnear Alters. Neural Computaon, Vol. 8-12(2000), p. 1743-1772. 4. Jannson T., Forrester T., and Degrood K. Wreless synapses n bo-nspred neural networks. Proc. of SPIE 09. Vol. 7374 (2009), p. 1-13. 5. Yan L.M. and Lu B. Optmal control of swtchng systems wth mpulsve effects. Internatonal Journal of Advances on dfferental equatons. Vol. 4 (2010), p. 1-14. 6. Tarek Oun,,Arj Lassoued, Mohamed Abd. Lossless mage compresson usng gradent based space fllng curves (G-SFC). Sgnal, Image and Vdeo Processng. Vol. 3 (2013), p. 1863-1703. 24