SVR_NL DCTs 'retransformadas' of a 2D signal subblock (Currently 256 * 256 resolution 4) Difference of QDCT63_2 in allowing the introduction of a mask (16 * 16 ) selecting certain coefficients so that, increased sequentially, you can simulate the progressive decoding (sending bits... : see progressive_coding.m ). This mask is introduced into the variable 'progres'. If progres = ones ( 16,16 ), the result is the same as using qdct63_2 Difference of QDCT62 that allows you to not calculate the eigenvalues of the original (if you're not interested, it saves time!) Represents QDCT62 signal in a domain with (approx.) statistically uncorrelated characteristics and uncorrelated perceptually (With perceptual metric identity). This is achieved through the divisive normalization: Teo & Heeger, Watson and Simoncelli. This transformation depends on its input signal ( state adaptation ). It has been shown that this type of response obtains independent statistics ( Simoncelli99 ) and perceptually ( Malo99 ) coefficients, and therefore is a suitable domain for a scalar quantizer. Since the perceptual metric in this domain is the identity apply uniform quantizers and flat bit-allocation per coefficient. For each block the invertibility is checked [max( eig ( D_R * h )) ], and the inverse is computed by using the analytical expression It is normalized to local Michaelson contrast and response according NAKA8.m: Norm = [1 1] ................ Normalized by the average luminance of the subfield considered (local contrast) . Norm = [3 Lm] ............... Normalized by the fixed value Lm (global contrast) the logical choice is to take Lm = mean ( mean ( Im) ); ( Mean value in digital levels ) Range to distribute the quantization levels in the domain response is given by the maximum diagonal gradient (acting on contrasts, belonging to [ -1.1 ]). USE: RESP=svr_nl_encoder(Im,bits,epsilon,K,exponente,directorio,fichero,RESP) Im -> Image to compress bits -> Numerb of bits for the SVR weights quantization epsilon -> Regression threshold K -> SVR Kernel directorio -> Place where the compressed file will be stored fichero -> name with which the compressed file will be stored RESP -> Optional parameter : nonlinear transform image In the first execution this parameter may not be available, but in a loop search of a certain entropy or distortion is calculated the first time and avoiding load posterior calculation.
0001 % SVR_NL DCTs 'retransformadas' of a 2D signal subblock 0002 % (Currently 256 * 256 resolution 4) 0003 % 0004 % Difference of QDCT63_2 in allowing the introduction of a mask 0005 % (16 * 16 ) selecting certain coefficients so that, 0006 % increased sequentially, you can simulate the progressive decoding 0007 % (sending bits... : see progressive_coding.m ). 0008 % This mask is introduced into the variable 'progres'. 0009 % If progres = ones ( 16,16 ), the result is the same as using qdct63_2 0010 % 0011 % Difference of QDCT62 that allows you to not calculate the eigenvalues of the original 0012 % (if you're not interested, it saves time!) 0013 % 0014 % Represents QDCT62 signal in a domain with (approx.) statistically uncorrelated characteristics 0015 % and uncorrelated perceptually (With perceptual metric identity). 0016 % 0017 % This is achieved through the divisive normalization: Teo & Heeger, Watson and Simoncelli. 0018 % This transformation depends on its input signal ( state adaptation ). 0019 % 0020 % It has been shown that this type of response obtains independent statistics ( Simoncelli99 ) and 0021 % perceptually ( Malo99 ) coefficients, and therefore is a suitable domain for 0022 % a scalar quantizer. 0023 % 0024 % Since the perceptual metric in this domain is the identity apply uniform quantizers 0025 % and flat bit-allocation per coefficient. 0026 % 0027 % For each block the invertibility is checked [max( eig ( D_R * h )) ], and the inverse is computed 0028 % by using the analytical expression 0029 % 0030 % It is normalized to local Michaelson contrast and response according NAKA8.m: 0031 % 0032 % Norm = [1 1] ................ Normalized by the average luminance 0033 % of the subfield considered (local contrast) . 0034 % 0035 % Norm = [3 Lm] ............... Normalized by the fixed value Lm (global contrast) 0036 % the logical choice is to take Lm = mean ( mean ( Im) ); 0037 % ( Mean value in digital levels ) 0038 % 0039 % Range to distribute the quantization levels in the domain response is given 0040 % by the maximum diagonal gradient (acting on contrasts, belonging to 0041 % [ -1.1 ]). 0042 % 0043 % USE: RESP=svr_nl_encoder(Im,bits,epsilon,K,exponente,directorio,fichero,RESP) 0044 % 0045 % Im -> Image to compress 0046 % bits -> Numerb of bits for the SVR weights quantization 0047 % epsilon -> Regression threshold 0048 % K -> SVR Kernel 0049 % directorio -> Place where the compressed file will be stored 0050 % fichero -> name with which the compressed file will be stored 0051 % RESP -> Optional parameter : nonlinear transform image 0052 % In the first execution this parameter may not be available, 0053 % but in a loop search of a certain entropy or 0054 % distortion is calculated the first time and avoiding load 0055 % posterior calculation. 0056 0057 function RESP=svr_nl_encoder(Im,bits,epsilon,K,exponente,directorio,fichero,RESP) 0058 0059 if nargin<8 0060 RESP=[]; 0061 end 0062 0063 MM = 0.8; 0064 normaliz = [3 mean(mean(Im))]; 0065 cero = 1; 0066 0067 [h,alpha,beta,gamm] = constrains_resp(exponente,cero); 0068 0069 Ncuan=4; 0070 tam=size(Im); 0071 tam=tam(1); 0072 lados=[tam tam]; 0073 lcuan=tam/(2^Ncuan); 0074 0075 posai=[tam tam]/2-round(tam/2); 0076 posbd=[tam tam]/2+round(tam/2); 0077 coorcuanai=floor([(posai(1)-1)/lcuan+1 (posai(2)-1)/lcuan+1]); 0078 coorcuanbd=floor([(posbd(1)-1)/lcuan+1 (posbd(2)-1)/lcuan+1]); 0079 if coorcuanai(1)<1 0080 coorcuanai(1)=1; 0081 end 0082 if coorcuanai(2)<1 0083 coorcuanai(2)=1; 0084 end 0085 if coorcuanbd(1)>(2^Ncuan) 0086 coorcuanbd(1)=2^Ncuan; 0087 end 0088 if coorcuanai(2)>(2^Ncuan) 0089 coorcuanbd(2)=2^Ncuan; 0090 end 0091 0092 dcti=dct2r(Im,Ncuan); 0093 0094 [dct_contr,Lm] = contras2(dcti,Ncuan,normaliz); 0095 DC = Lm*16; 0096 0097 W=1; 0098 0099 val_r = 100; 0100 0101 e = imdpcm(Lm,W,val_r); 0102 0103 N = round(0.9 * 40); 0104 0105 me = mini(e); 0106 Me = maxi(e); 0107 0108 extr_err = [me Me]; 0109 0110 QE = round((N -1) * (e - me) / (Me - me)); 0111 0112 R=(1:255);R=R(:); 0113 0114 C=40000; 0115 perfil_C=ones(255,1); 0116 0117 perfil_e=ones(255,1); 0118 0119 Lb = lcuan; 0120 Li = tam; 0121 0122 Results=[]; 0123 0124 codigazo = []; 0125 0126 signosr = zeros(256,255); 0127 betasc = zeros(256,255); 0128 betascr = zeros(256,255); 0129 0130 if nargin<8 0131 fprintf(' Computing the non-linear response \n'); 0132 fprintf(' SVM learning the blocks of the local non-linear domain'); 0133 else 0134 fprintf(' SVM learning the blocks of the local non-linear domain'); 0135 end 0136 0137 bloquecito = 1; 0138 for i=1:Lb:Li 0139 0140 fprintf('.'); 0141 0142 for j=1:Lb:Li 0143 0144 aa = dct_contr(i:(i+(Lb-1)),j:(j+(Lb-1))); 0145 0146 if nargin<8 0147 [h_a,r,GR,alfa2d,beta2d] = respue5(aa,alpha,beta,gamm,h,0); 0148 0149 r=zigzag(r); 0150 0151 RESP=[RESP r]; 0152 else 0153 r=RESP(:,bloquecito); 0154 end 0155 0156 [nsv,pesosSVM,bias] = irwls_pd_svr_nobias(R,r(2:256),K,C,perfil_C,epsilon,perfil_e); 0157 0158 azig = zigzag(aa); 0159 azigsc = azig(2:end); 0160 0161 bitsig = getsign(azigsc,Lb); 0162 0163 signosr(bloquecito,:) = bitsig; 0164 betasc(bloquecito,:) = pesosSVM'; 0165 0166 bloquecito = bloquecito + 1; 0167 0168 end 0169 end 0170 fprintf('\n'); 0171 0172 NumNiv = 2^bits; 0173 0174 betascr = round( NumNiv * betasc / MM) * MM / NumNiv; 0175 0176 rc = []; 0177 for fila = 1:Li 0178 rc = [rc; (K*betascr(fila,:)')']; 0179 end 0180 0181 signos = zeros(size(signosr)); 0182 signos(logical(abs(rc)>0))=signosr(logical(abs(rc)>0)); 0183 0184 codigon = betascr.*(~signos) - betascr.*(signos); 0185 0186 codigonint = codigon * NumNiv / MM; 0187 0188 for i=1:Lb^2 0189 0190 [cod,ceros] = rle2(codigonint(i,:)); 0191 0192 codigazo = [codigazo cod ceros]; 0193 0194 end 0195 0196 ristra = [mean(mean(Im)) extr_err(:)' round(QE(:))' bits round(codigazo)]; 0197 0198 cd(directorio); 0199 save_code_svr_nl(ristra,[fichero '.bin']); 0200 0201 zip(fichero,[fichero '.bin']); 0202 0203 delete([fichero '.bin']); 0204