001/* 002 * $RCSfile: StdQuantizer.java,v $ 003 * $Revision: 1.1 $ 004 * $Date: 2005/02/11 05:02:20 $ 005 * $State: Exp $ 006 * 007 * Class: StdQuantizer 008 * 009 * Description: Scalar deadzone quantizer of integer or float 010 * data. 011 * 012 * Mergerd from StdQuantizerInt and 013 * StdQuantizerFloat from Joel Askelof. 014 * 015 * 016 * COPYRIGHT: 017 * 018 * This software module was originally developed by Raphaël Grosbois and 019 * Diego Santa Cruz (Swiss Federal Institute of Technology-EPFL); Joel 020 * Askelöf (Ericsson Radio Systems AB); and Bertrand Berthelot, David 021 * Bouchard, Félix Henry, Gerard Mozelle and Patrice Onno (Canon Research 022 * Centre France S.A) in the course of development of the JPEG2000 023 * standard as specified by ISO/IEC 15444 (JPEG 2000 Standard). This 024 * software module is an implementation of a part of the JPEG 2000 025 * Standard. Swiss Federal Institute of Technology-EPFL, Ericsson Radio 026 * Systems AB and Canon Research Centre France S.A (collectively JJ2000 027 * Partners) agree not to assert against ISO/IEC and users of the JPEG 028 * 2000 Standard (Users) any of their rights under the copyright, not 029 * including other intellectual property rights, for this software module 030 * with respect to the usage by ISO/IEC and Users of this software module 031 * or modifications thereof for use in hardware or software products 032 * claiming conformance to the JPEG 2000 Standard. Those intending to use 033 * this software module in hardware or software products are advised that 034 * their use may infringe existing patents. The original developers of 035 * this software module, JJ2000 Partners and ISO/IEC assume no liability 036 * for use of this software module or modifications thereof. No license 037 * or right to this software module is granted for non JPEG 2000 Standard 038 * conforming products. JJ2000 Partners have full right to use this 039 * software module for his/her own purpose, assign or donate this 040 * software module to any third party and to inhibit third parties from 041 * using this software module for non JPEG 2000 Standard conforming 042 * products. This copyright notice must be included in all copies or 043 * derivative works of this software module. 044 * 045 * Copyright (c) 1999/2000 JJ2000 Partners. 046 * */ 047package jj2000.j2k.quantization.quantizer; 048import jj2000.j2k.codestream.writer.*; 049import jj2000.j2k.wavelet.analysis.*; 050import jj2000.j2k.quantization.*; 051import jj2000.j2k.wavelet.*; 052import jj2000.j2k.image.*; 053import jj2000.j2k.*; 054 055import com.sun.media.imageioimpl.plugins.jpeg2000.J2KImageWriteParamJava; 056 057/** 058 * This class implements scalar quantization of integer or floating-point 059 * valued source data. The source data is the wavelet transformed image data 060 * and the output is the quantized wavelet coefficients represented in 061 * sign-magnitude (see below). 062 * 063 * <P>Sign magnitude representation is used (instead of two's complement) for 064 * the output data. The most significant bit is used for the sign (0 if 065 * positive, 1 if negative). Then the magnitude of the quantized coefficient 066 * is stored in the next M most significat bits. The rest of the bits (least 067 * significant bits) can contain a fractional value of the quantized 068 * coefficient. This fractional value is not to be coded by the entropy 069 * coder. However, it can be used to compute rate-distortion measures with 070 * greater precision. 071 * 072 * <P>The value of M is determined for each subband as the sum of the number 073 * of guard bits G and the nominal range of quantized wavelet coefficients in 074 * the corresponding subband (Rq), minus 1: 075 * 076 * <P>M = G + Rq -1 077 * 078 * <P>The value of G should be the same for all subbands. The value of Rq 079 * depends on the quantization step size, the nominal range of the component 080 * before the wavelet transform and the analysis gain of the subband (see 081 * Subband). 082 * 083 * <P>The blocks of data that are requested should not cross subband 084 * boundaries. 085 * 086 * @see Subband 087 * 088 * @see Quantizer 089 * */ 090public class StdQuantizer extends Quantizer { 091 092 /** The number of mantissa bits for the quantization steps */ 093 public final static int QSTEP_MANTISSA_BITS = 11; 094 095 /** The number of exponent bits for the quantization steps */ 096 // NOTE: formulas in 'convertFromExpMantissa()' and 097 // 'convertToExpMantissa()' methods do not support more than 5 bits. 098 public final static int QSTEP_EXPONENT_BITS = 5; 099 100 /** The maximum value of the mantissa for the quantization steps */ 101 public final static int QSTEP_MAX_MANTISSA = (1<<QSTEP_MANTISSA_BITS)-1; 102 103 /** The maximum value of the exponent for the quantization steps */ 104 public final static int QSTEP_MAX_EXPONENT = (1<<QSTEP_EXPONENT_BITS)-1; 105 106 /** Natural log of 2, used as a convenience variable */ 107 private static double log2 = Math.log(2); 108 109 /** The quantization type specifications */ 110 private QuantTypeSpec qts; 111 112 /** The quantization step size specifications */ 113 private QuantStepSizeSpec qsss; 114 115 /** The guard bits specifications */ 116 private GuardBitsSpec gbs; 117 118 /** The 'CBlkWTDataFloat' object used to request data, used when 119 * quantizing floating-point data. */ 120 // This variable makes the class thread unsafe, but it avoids allocating 121 // new objects for code-block that is quantized. 122 private CBlkWTDataFloat infblk; 123 124 /** 125 * Initializes the source of wavelet transform coefficients. The 126 * constructor takes information on whether the quantizer is in 127 * reversible, derived or expounded mode. If the quantizer is reversible 128 * the value of 'derived' is ignored. If the source data is not integer 129 * (int) then the quantizer can not be reversible. 130 * 131 * <P> After initializing member attributes, getAnSubbandTree is called for 132 * all components setting the 'stepWMSE' for all subbands in the current 133 * tile. 134 * 135 * @param src The source of wavelet transform coefficients. 136 * 137 * @param encSpec The encoder specifications 138 * */ 139 public StdQuantizer(CBlkWTDataSrc src,J2KImageWriteParamJava wp){ 140 super(src); 141 qts = wp.getQuantizationType(); 142 qsss = wp.getQuantizationStep(); 143 gbs = wp.getGuardBits(); 144 } 145 146 /** 147 * Returns the quantization type spec object associated to the quantizer. 148 * 149 * @return The quantization type spec 150 * */ 151 public QuantTypeSpec getQuantTypeSpec(){ 152 return qts; 153 } 154 155 /** 156 * Returns the number of guard bits used by this quantizer in the given 157 * tile-component. 158 * 159 * @param t Tile index 160 * 161 * @param c Component index 162 * 163 * @return The number of guard bits 164 * */ 165 public int getNumGuardBits(int t,int c){ 166 return ((Integer)gbs.getTileCompVal(t,c)).intValue(); 167 } 168 169 /** 170 * Returns true if the quantized data is reversible, for the specified 171 * tile-component. For the quantized data to be reversible it is necessary 172 * and sufficient that the quantization is reversible. 173 * 174 * @param t The tile to test for reversibility 175 * 176 * @param c The component to test for reversibility 177 * 178 * @return True if the quantized data is reversible, false if not. 179 * */ 180 public boolean isReversible(int t,int c){ 181 return qts.isReversible(t,c); 182 } 183 184 /** 185 * Returns true if given tile-component uses derived quantization step 186 * sizes. 187 * 188 * @param t Tile index 189 * 190 * @param c Component index 191 * 192 * @return True if derived 193 * 194 */ 195 public boolean isDerived(int t,int c){ 196 return qts.isDerived(t,c); 197 } 198 199 /** 200 * Returns the next code-block in the current tile for the specified 201 * component, as a copy (see below). The order in which code-blocks are 202 * returned is not specified. However each code-block is returned only 203 * once and all code-blocks will be returned if the method is called 'N' 204 * times, where 'N' is the number of code-blocks in the tile. After all 205 * the code-blocks have been returned for the current tile calls to this 206 * method will return 'null'. 207 * 208 * <P>When changing the current tile (through 'setTile()' or 'nextTile()') 209 * this method will always return the first code-block, as if this method 210 * was never called before for the new current tile. 211 * 212 * <P>The data returned by this method is always a copy of the 213 * data. Therfore it can be modified "in place" without any problems after 214 * being returned. The 'offset' of the returned data is 0, and the 'scanw' 215 * is the same as the code-block width. See the 'CBlkWTData' class. 216 * 217 * <P>The 'ulx' and 'uly' members of the returned 'CBlkWTData' object 218 * contain the coordinates of the top-left corner of the block, with 219 * respect to the tile, not the subband. 220 * 221 * @param c The component for which to return the next code-block. 222 * 223 * @param cblk If non-null this object will be used to return the new 224 * code-block. If null a new one will be allocated and returned. If the 225 * "data" array of the object is non-null it will be reused, if possible, 226 * to return the data. 227 * 228 * @return The next code-block in the current tile for component 'n', or 229 * null if all code-blocks for the current tile have been returned. 230 * 231 * @see CBlkWTData 232 * */ 233 public CBlkWTData getNextCodeBlock(int c,CBlkWTData cblk) { 234 return getNextInternCodeBlock(c,cblk); 235 } 236 237 /** 238 * Returns the next code-block in the current tile for the specified 239 * component. The order in which code-blocks are returned is not 240 * specified. However each code-block is returned only once and all 241 * code-blocks will be returned if the method is called 'N' times, where 242 * 'N' is the number of code-blocks in the tile. After all the code-blocks 243 * have been returned for the current tile calls to this method will 244 * return 'null'. 245 * 246 * <P>When changing the current tile (through 'setTile()' or 'nextTile()') 247 * this method will always return the first code-block, as if this method 248 * was never called before for the new current tile. 249 * 250 * <P>The data returned by this method can be the data in the internal 251 * buffer of this object, if any, and thus can not be modified by the 252 * caller. The 'offset' and 'scanw' of the returned data can be 253 * arbitrary. See the 'CBlkWTData' class. 254 * 255 * <P>The 'ulx' and 'uly' members of the returned 'CBlkWTData' object 256 * contain the coordinates of the top-left corner of the block, with 257 * respect to the tile, not the subband. 258 * 259 * @param c The component for which to return the next code-block. 260 * 261 * @param cblk If non-null this object will be used to return the new 262 * code-block. If null a new one will be allocated and returned. If the 263 * "data" array of the object is non-null it will be reused, if possible, 264 * to return the data. 265 * 266 * @return The next code-block in the current tile for component 'n', or 267 * null if all code-blocks for the current tile have been returned. 268 * 269 * @see CBlkWTData 270 * */ 271 public final CBlkWTData getNextInternCodeBlock(int c, CBlkWTData cblk) { 272 // NOTE: this method is declared final since getNextCodeBlock() relies 273 // on this particular implementation 274 int k,j; 275 int tmp,shiftBits,jmin; 276 int w,h; 277 int outarr[]; 278 float infarr[] = null; 279 CBlkWTDataFloat infblk; 280 float invstep; // The inverse of the quantization step size 281 boolean intq; // flag for quantizig ints 282 SubbandAn sb; 283 float stepUDR; // The quantization step size (for a dynamic 284 // range of 1, or unit) 285 int g = ((Integer)gbs.getTileCompVal(tIdx,c)).intValue(); 286 287 // Are we quantizing ints or floats? 288 intq = (src.getDataType(tIdx,c) == DataBlk.TYPE_INT); 289 290 // Check that we have an output object 291 if (cblk == null) { 292 cblk = new CBlkWTDataInt(); 293 } 294 295 // Cache input float code-block 296 infblk = this.infblk; 297 298 // Get data to quantize. When quantizing int data 'cblk' is used to 299 // get the data to quantize and to return the quantized data as well, 300 // that's why 'getNextCodeBlock()' is used. This can not be done when 301 // quantizing float data because of the different data types, that's 302 // why 'getNextInternCodeBlock()' is used in that case. 303 if (intq) { // Source data is int 304 cblk = src.getNextCodeBlock(c,cblk); 305 if (cblk == null) { 306 return null; // No more code-blocks in current tile for comp. 307 } 308 // Input and output arrays are the same (for "in place" quant.) 309 outarr = (int[])cblk.getData(); 310 } 311 else { // Source data is float 312 // Can not use 'cblk' to get float data, use 'infblk' 313 infblk = (CBlkWTDataFloat) src.getNextInternCodeBlock(c,infblk); 314 if (infblk == null) { 315 // Release buffer from infblk: this enables to garbage collect 316 // the big buffer when we are done with last code-block of 317 // component. 318 this.infblk.setData(null); 319 return null; // No more code-blocks in current tile for comp. 320 } 321 this.infblk = infblk; // Save local cache 322 infarr = (float[])infblk.getData(); 323 // Get output data array and check that there is memory to put the 324 // quantized coeffs in 325 outarr = (int[]) cblk.getData(); 326 if (outarr == null || outarr.length < infblk.w*infblk.h) { 327 outarr = new int[infblk.w*infblk.h]; 328 cblk.setData(outarr); 329 } 330 cblk.m = infblk.m; 331 cblk.n = infblk.n; 332 cblk.sb = infblk.sb; 333 cblk.ulx = infblk.ulx; 334 cblk.uly = infblk.uly; 335 cblk.w = infblk.w; 336 cblk.h = infblk.h; 337 cblk.wmseScaling = infblk.wmseScaling; 338 cblk.offset = 0; 339 cblk.scanw = cblk.w; 340 } 341 342 // Cache width, height and subband of code-block 343 w = cblk.w; 344 h = cblk.h; 345 sb = cblk.sb; 346 347 if(isReversible(tIdx,c)) { // Reversible only for int data 348 cblk.magbits = g-1+src.getNomRangeBits(c)+sb.anGainExp; 349 shiftBits = 31-cblk.magbits; 350 351 // Update the convertFactor field 352 cblk.convertFactor = (1<<shiftBits); 353 354 // Since we used getNextCodeBlock() to get the int data then 355 // 'offset' is 0 and 'scanw' is the width of the code-block The 356 // input and output arrays are the same (i.e. "in place") 357 for(j=w*h-1; j>=0; j--){ 358 tmp = (outarr[j]<<shiftBits); 359 outarr[j] = ((tmp < 0) ? (1<<31)|(-tmp) : tmp); 360 } 361 } 362 else{ // Non-reversible, use step size 363 float baseStep = 364 ((Float)qsss.getTileCompVal(tIdx,c)).floatValue(); 365 366 // Calculate magnitude bits and quantization step size 367 if(isDerived(tIdx,c)){ 368 cblk.magbits = g-1+sb.level- 369 (int)Math.floor(Math.log(baseStep)/log2); 370 stepUDR = baseStep/(1<<sb.level); 371 } 372 else{ 373 cblk.magbits = g-1-(int)Math.floor(Math.log(baseStep/ 374 (sb.l2Norm*(1<<sb.anGainExp)))/ 375 log2); 376 stepUDR = baseStep/(sb.l2Norm*(1<<sb.anGainExp)); 377 } 378 shiftBits = 31-cblk.magbits; 379 // Calculate step that decoder will get and use that one. 380 stepUDR = 381 convertFromExpMantissa(convertToExpMantissa(stepUDR)); 382 invstep = 1.0f/((1L<<(src.getNomRangeBits(c)+sb.anGainExp))* 383 stepUDR); 384 // Normalize to magnitude bits (output fractional point) 385 invstep *= (1<<(shiftBits-src.getFixedPoint(c))); 386 387 // Update convertFactor and stepSize fields 388 cblk.convertFactor = invstep; 389 cblk.stepSize = ((1L<<(src.getNomRangeBits(c)+sb.anGainExp))* 390 stepUDR); 391 392 if(intq){ // Quantizing int data 393 // Since we used getNextCodeBlock() to get the int data then 394 // 'offset' is 0 and 'scanw' is the width of the code-block 395 // The input and output arrays are the same (i.e. "in place") 396 for (j=w*h-1; j>=0; j--) { 397 tmp = (int)(outarr[j]*invstep); 398 outarr[j] = ((tmp < 0) ? (1<<31)|(-tmp) : tmp); 399 } 400 } 401 else { // Quantizing float data 402 for (j=w*h-1, k = infblk.offset+(h-1)*infblk.scanw+w-1, 403 jmin = w*(h-1); j>=0; jmin -= w) { 404 for (; j>=jmin; k--, j--) { 405 tmp = (int)(infarr[k]*invstep); 406 outarr[j] = ((tmp < 0) ? (1<<31)|(-tmp) : tmp); 407 } 408 // Jump to beggining of previous line in input 409 k -= infblk.scanw - w; 410 } 411 } 412 } 413 // Return the quantized code-block 414 return cblk; 415 } 416 417 /** 418 * Calculates the parameters of the SubbandAn objects that depend on the 419 * Quantizer. The 'stepWMSE' field is calculated for each subband which is 420 * a leaf in the tree rooted at 'sb', for the specified component. The 421 * subband tree 'sb' must be the one for the component 'n'. 422 * 423 * @param sb The root of the subband tree. 424 * 425 * @param c The component index 426 * 427 * @see SubbandAn#stepWMSE 428 * */ 429 protected void calcSbParams(SubbandAn sb,int c){ 430 float baseStep; 431 432 if(sb.stepWMSE>0f) // parameters already calculated 433 return; 434 if(!sb.isNode){ 435 if(isReversible(tIdx,c)){ 436 sb.stepWMSE = (float) Math.pow(2,-(src.getNomRangeBits(c)<<1))* 437 sb.l2Norm*sb.l2Norm; 438 } 439 else{ 440 baseStep = ((Float)qsss.getTileCompVal(tIdx,c)).floatValue(); 441 if(isDerived(tIdx,c)){ 442 sb.stepWMSE = baseStep*baseStep* 443 (float)Math.pow(2,(sb.anGainExp-sb.level)<<1)* 444 sb.l2Norm*sb.l2Norm; 445 } 446 else{ 447 sb.stepWMSE = baseStep*baseStep; 448 } 449 } 450 } 451 else{ 452 calcSbParams((SubbandAn)sb.getLL(),c); 453 calcSbParams((SubbandAn)sb.getHL(),c); 454 calcSbParams((SubbandAn)sb.getLH(),c); 455 calcSbParams((SubbandAn)sb.getHH(),c); 456 sb.stepWMSE = 1f; // Signal that we already calculated this branch 457 } 458 } 459 460 /** 461 * Converts the floating point value to its exponent-mantissa 462 * representation. The mantissa occupies the 11 least significant bits 463 * (bits 10-0), and the exponent the previous 5 bits (bits 15-11). 464 * 465 * @param step The quantization step, normalized to a dynamic range of 1. 466 * 467 * @return The exponent mantissa representation of the step. 468 * */ 469 public static int convertToExpMantissa(float step) { 470 int exp; 471 472 exp = (int)Math.ceil(-Math.log(step)/log2); 473 if (exp>QSTEP_MAX_EXPONENT) { 474 // If step size is too small for exponent representation, use the 475 // minimum, which is exponent QSTEP_MAX_EXPONENT and mantissa 0. 476 return (QSTEP_MAX_EXPONENT<<QSTEP_MANTISSA_BITS); 477 } 478 // NOTE: this formula does not support more than 5 bits for the 479 // exponent, otherwise (-1<<exp) might overflow (the - is used to be 480 // able to represent 2**31) 481 return (exp<<QSTEP_MANTISSA_BITS) | 482 ((int)((-step*(-1<<exp)-1f)*(1<<QSTEP_MANTISSA_BITS)+0.5f)); 483 } 484 485 /** 486 * Converts the exponent-mantissa representation to its floating-point 487 * value. The mantissa occupies the 11 least significant bits (bits 10-0), 488 * and the exponent the previous 5 bits (bits 15-11). 489 * 490 * @param ems The exponent-mantissa representation of the step. 491 * 492 * @return The floating point representation of the step, normalized to a 493 * dynamic range of 1. 494 * */ 495 private static float convertFromExpMantissa(int ems) { 496 // NOTE: this formula does not support more than 5 bits for the 497 // exponent, otherwise (-1<<exp) might overflow (the - is used to be 498 // able to represent 2**31) 499 return (-1f-((float)(ems&QSTEP_MAX_MANTISSA)) / 500 ((float)(1<<QSTEP_MANTISSA_BITS))) / 501 (float)(-1<<((ems>>QSTEP_MANTISSA_BITS)&QSTEP_MAX_EXPONENT)); 502 } 503 504 /** 505 * Returns the maximum number of magnitude bits in any subband of the 506 * current tile. 507 * 508 * @param c the component number 509 * 510 * @return The maximum number of magnitude bits in all subbands of the 511 * current tile. 512 * */ 513 public int getMaxMagBits(int c){ 514 Subband sb = getAnSubbandTree(tIdx,c); 515 if(isReversible(tIdx,c)){ 516 return getMaxMagBitsRev(sb,c); 517 } 518 else{ 519 if(isDerived(tIdx,c)){ 520 return getMaxMagBitsDerived(sb,tIdx,c); 521 } 522 else { 523 return getMaxMagBitsExpounded(sb,tIdx,c); 524 } 525 } 526 } 527 528 529 /** 530 * Returns the maximum number of magnitude bits in any subband of the 531 * current tile if reversible quantization is used 532 * 533 * @param sb The root of the subband tree of the current tile 534 * 535 * @param c the component number 536 * 537 * @return The highest number of magnitude bit-planes 538 * */ 539 private int getMaxMagBitsRev(Subband sb, int c){ 540 int tmp,max=0; 541 int g = ((Integer)gbs.getTileCompVal(tIdx,c)).intValue(); 542 543 if(!sb.isNode) 544 return g-1+src.getNomRangeBits(c)+sb.anGainExp; 545 546 max=getMaxMagBitsRev(sb.getLL(),c); 547 tmp=getMaxMagBitsRev(sb.getLH(),c); 548 if(tmp>max) 549 max=tmp; 550 tmp=getMaxMagBitsRev(sb.getHL(),c); 551 if(tmp>max) 552 max=tmp; 553 tmp=getMaxMagBitsRev(sb.getHH(),c); 554 if(tmp>max) 555 max=tmp; 556 557 return max; 558 } 559 560 /** 561 * Returns the maximum number of magnitude bits in any subband in the 562 * given tile-component if derived quantization is used 563 * 564 * @param sb The root of the subband tree of the tile-component 565 * 566 * @param t Tile index 567 * 568 * @param c Component index 569 * 570 * @return The highest number of magnitude bit-planes 571 * */ 572 private int getMaxMagBitsDerived(Subband sb,int t,int c){ 573 int tmp,max=0; 574 int g = ((Integer)gbs.getTileCompVal(t,c)).intValue(); 575 576 if(!sb.isNode){ 577 float baseStep = ((Float)qsss.getTileCompVal(t,c)).floatValue(); 578 return g-1+sb.level-(int)Math.floor(Math.log(baseStep)/log2); 579 } 580 581 max=getMaxMagBitsDerived(sb.getLL(),t,c); 582 tmp=getMaxMagBitsDerived(sb.getLH(),t,c); 583 if(tmp>max) 584 max=tmp; 585 tmp=getMaxMagBitsDerived(sb.getHL(),t,c); 586 if(tmp>max) 587 max=tmp; 588 tmp=getMaxMagBitsDerived(sb.getHH(),t,c); 589 if(tmp>max) 590 max=tmp; 591 592 return max; 593 } 594 595 596 /** 597 * Returns the maximum number of magnitude bits in any subband in the 598 * given tile-component if expounded quantization is used 599 * 600 * @param sb The root of the subband tree of the tile-component 601 * 602 * @param t Tile index 603 * 604 * @param c Component index 605 * 606 * @return The highest number of magnitude bit-planes 607 * */ 608 private int getMaxMagBitsExpounded(Subband sb,int t,int c){ 609 int tmp,max=0; 610 int g = ((Integer)gbs.getTileCompVal(t,c)).intValue(); 611 612 if(!sb.isNode){ 613 float baseStep = ((Float)qsss.getTileCompVal(t,c)).floatValue(); 614 return g-1- 615 (int)Math.floor(Math.log(baseStep/ 616 (((SubbandAn)sb).l2Norm*(1<<sb.anGainExp)))/ 617 log2); 618 } 619 620 max=getMaxMagBitsExpounded(sb.getLL(),t,c); 621 tmp=getMaxMagBitsExpounded(sb.getLH(),t,c); 622 if(tmp>max) 623 max=tmp; 624 tmp=getMaxMagBitsExpounded(sb.getHL(),t,c); 625 if(tmp>max) 626 max=tmp; 627 tmp=getMaxMagBitsExpounded(sb.getHH(),t,c); 628 if(tmp>max) 629 max=tmp; 630 631 return max; 632 } 633}