矢量量化實(shí)驗(yàn)報告

1、 矢量量化與圖像處理實(shí)驗(yàn)報告 設(shè)計(jì)課題:矢量量化的碼書設(shè)計(jì)技術(shù) 姓 名: 學(xué) 院: 專 業(yè):班 級: 學(xué) 號:指導(dǎo)教師: 實(shí)驗(yàn)一：矢量量化的碼書設(shè)計(jì)技術(shù) 一、實(shí)驗(yàn)?zāi)康模?1. 掌握矢量量化技術(shù)中碼書設(shè)計(jì)的基本原理； 2. 掌握經(jīng)典的LBG算法； 3. 了解初始碼書選取在LBG算法中的重要性； 4. 掌握基于神經(jīng)網(wǎng)絡(luò)和遺傳算法等優(yōu)化算法的碼書設(shè)計(jì)算法； 5. 嘗試設(shè)計(jì)基于LBG的碼書生成算法5-10（從提高碼書質(zhì)量和加速訓(xùn)練速度等方面出發(fā)）； 6. 嘗試設(shè)計(jì)基于優(yōu)化算法的碼書生成算法11-13。二、實(shí)驗(yàn)內(nèi)容： 1. 基礎(chǔ)部分： 完成以下任意兩種初始碼書算法并給出實(shí)驗(yàn)結(jié)果： 1）隨機(jī)法初始碼書算法

2、，算法描述詳見參考文獻(xiàn)1-3； 2）分離平均初始碼書算法，算法描述詳見參考文獻(xiàn)1, 4； 3）競爭學(xué)習(xí)矢量量化算法，算法描述詳見參考文獻(xiàn)1-3, 10。2. 提高部分： 任選一個或者多個矢量特征量（均值、方差、范數(shù)等）設(shè)計(jì)基于矢量特征量排序的分離平均初始碼書算法，算法描述詳見參考文獻(xiàn)1, 6, 8, 9。3. 發(fā)揮部分： 改進(jìn)文獻(xiàn)5-13中的算法或者設(shè)計(jì)新的碼書生成算法并詳細(xì)給出實(shí)驗(yàn)設(shè)計(jì)的方案、實(shí)驗(yàn)結(jié)果以及實(shí)驗(yàn)結(jié)論。 三、實(shí)驗(yàn)要求： 1. 用單幅圖像和多幅圖像的組合作為矢量訓(xùn)練集， 2. 至少分別用Lena、Peppers、Airplane和Baboon四幅圖像以及其組合圖像作為訓(xùn)練集圖像（I

3、nside Images），不在訓(xùn)練集內(nèi)的圖像作為訓(xùn)練集外的編碼圖像（Outside Images），用訓(xùn)練集圖像和訓(xùn)練集外的圖像對碼書進(jìn)行測試，測試算法的有效性和穩(wěn)健性，圖像大小均為512512的8-bit灰度圖像； ×3. 矢量維數(shù)為16維（44的圖像塊構(gòu)成）； ×4. 碼書尺寸分別為64、128、256、512和1024； 5. 主要實(shí)驗(yàn)數(shù)據(jù)： 1）初始碼書對訓(xùn)練集圖像的編碼質(zhì)量，編碼質(zhì)量可以用峰值信噪比（PSNR）評估； 2）最終碼書對測試圖像（訓(xùn)練集和訓(xùn)練集外圖像）的編碼質(zhì)量； 3）碼書訓(xùn)練過程中迭代誤差與迭代次數(shù)關(guān)系圖，迭代誤差可以用均方誤差（MSE）評估； 4

4、）設(shè)計(jì)提高碼書訓(xùn)練速度的算法需給出碼書訓(xùn)練的時間或計(jì)算復(fù)雜度分析。 四、實(shí)驗(yàn)環(huán)境： C語言和Matlab均可，建議用C語言編程。實(shí)驗(yàn)結(jié)果：分別對碼書大小為64,128,256,516,1024進(jìn)行實(shí)驗(yàn) 表1.碼書大小為64隨機(jī)碼書誤差次數(shù)LenaAirplanebaboonpeppersLe+ AiLe+ Ai+peLe+ Ai+pe+ba訓(xùn)練MSE1100.8870181.6048368.0803110.456121.3566093127.2391141175.5118103275.8178102.9900303.510280.231583.8708719285.36996106142.40

5、21172370.575686.3671290.086371.279476.1858846577.86092818135.0865874467.370078.9570283.815766.621673.358112874.95641166131.3985009565.136774.2382280.133964.492271.5160129673.34756348129.2097119663.707071.0436277.513163.302470.2019896772.12705039127.9775801762.746769.0037275.554662.456169.3197736171.

6、07881048127.2000149862.156567.5102273.923461.945868.6466795870.35909907126.6103904961.737366.3061272.771461.655868.0405841469.88538401126.1242321061.416365.6286271.812461.388467.4624801469.48534731125.7021381161.156465.1921271.114361.103566.8766383469.1549258125.3645521260.957964.7903270.574760.8645

7、66.3979484868.9105475125.11482891360.803864.4687270.137660.612066.0628468268.71579676124.94191331460.694964.1689269.737060.332165.8456646668.53327249124.82221431560.568163.8838269.381960.078365.7315023968.38721895124.74497861660.443663.6169269.059159.898365.6476840468.25992385124.68184211760.346363.

8、3810268.801559.778265.559847066864169421860.268863.1832268.570259.701765.4820469768.02121164124.60968451960.190863.0232268.340859.629565.4407667.92175817124.57662060.110162.8346268.136359.571365.4093661567.84226636124.5472817PSNR128.0925 25.539522.471427.698927.2901727.084625.68774229.3

9、331 28.002923.309129.087428.8946928.8177526.59564329.6443 28.767323.505529.601229.3120629.2176126.82468429.8461 29.156923.600429.894729.4763229.3827226.9449529.9925 29.424523.657130.035729.5867729.4769527.01785630.0889 29.615623.815030.116629.6673129.5498227.05946730.154929.742123.728730.175129.7222

10、329.613427.08593830.195929.837123.754530.210729.7646129.657627.10611930.225329.915323.772830.231129.8031229.6869427.122821030.248029.959923.788130.249929.8401829.7118727.137381130.266429.988923.799330.270129.8780629.7325727.149061230.280530.015723.807930.287229.9092629.7479527.157721330.291530.03732

11、3.815030.305229.9312329.7602427.163721430.299330.057623.821430.325329.9455329.7717927.167881530.308430.076923.827130.343629.9530729.7810527.170571630.317330.095123.832330.356729.9586129.7891527.172771730.324330.111223.836530.365429.9644229.7970827.174171830.329930.124823.840230.370929.9695829.804362

12、7.175291930.335530.135823.843930.376229.9723229.8107127.176442030.341330.148823.847230.380429.974429.815827.17746測試PSNR126.466326.001025.479426.815625.932526.664126.9653通過表格的數(shù)據(jù)可以得知初始碼書對訓(xùn)練集圖像的編碼質(zhì)量，隨著迭代次數(shù)增加編碼質(zhì)量質(zhì)量逐漸提高在單幅圖像的作為訓(xùn)練集時，選擇圖像的質(zhì)量對均方誤差的影響，四幅圖像分別作為訓(xùn)練集時的MSE，如下圖所示： 不同圖像對MSE影響由以上的折線圖可以得出質(zhì)量比較差的圖片對MSE

13、的影響是很大的，例如圖片baboon512.bmp，這張圖片的效果是最差的，原因是我們希望產(chǎn)生的碼書盡可能包含有不同的圖像塊（4*4），狒狒這張圖像近似的圖像塊有很多，而且在圖片上這些圖像塊是呈對稱分布的，所以可能左邊的圖像塊可能被量化到右邊的碼字，或者右邊的圖像塊可能被量化到左邊的碼字，這都會造成很大的誤差。碼書訓(xùn)練過程中迭代誤差與迭代次數(shù)關(guān)系圖如下所示，迭代誤差用均方誤差（MSE）表示 迭代誤差與迭代次數(shù)關(guān)系圖有上圖可以看出隨著迭代次數(shù)的增加，MSE下降的速度均是越越來越慢，到最后基本上不會再下降。但是baboon的編碼質(zhì)量依然是最差的。同樣可以得出PSNR的規(guī)律：隨著訓(xùn)練集的增加，在迭代

14、的時候基本上呈現(xiàn)下降趨勢的，也就是說單幅圖迭代20次得出的PSNR比多副圖進(jìn)行訓(xùn)練迭代得出的PSNR要小。測試碼書的大?。?4,128,256,512,1024）對量化的影響，利用隨機(jī)法產(chǎn)生初始碼書，在進(jìn)行20次迭代，訓(xùn)練集時四副圖像，測試集是全部12張圖片。誤差次數(shù)641282565121024訓(xùn)練MSE1175.5118103226.5115623196.2291927158.288800296.185857772142.4021172143.6144456118.2689645105.209757975.815087273135.0865874133.5692544107.6630894

15、.5460320770.238819484131.3985009127.8655584102.360088788.4513723567.439101795129.2097119123.078833499.1021223184.310950765.747692836127.9775801118.561256896.9044179181.9729502964.630033587127.2000149114.620598195.2416065180.6116149963.851912448126.6103904111.648572593.8906180179.6293119963.229919659

16、126.124232110.072773192.8378526478.8764389762.7250502210125.702138109.150873891.9928931378.2845083162.2873916511125.364552108.473097691.3395364277.7709037261.9496446712125.1148289107.941336290.8164221777.3645309861.6701154213124.9419133107.527337990.4309705577.0423957261.4423345114124.8222143107.193

17、880590.1650752476.7631881361.2378552715124.7449786106.883641689.9379323476.5052834361.0673094716124.6818421106.615231189.735620476.2937096160.9124189417124.6416942106.387448689.5459439876.0995605660.780280518124.6096845106.199852289.3658005775.9162432660.6727939819124.5766434106.032904989.2080570675

18、.7481076360.5742901220124.5472817105.872410989.0766513575.5824328760.47858524PSNR125.6877401524.5798998526.1363017426.1363017428.2997226.5956391526.5588223527.910243427.910243429.33324722326.824681326.8737385928.3743705428.3743705429.66503157426.944899527.0632678128.6637578528.6637578529.84168584527

19、.0178520327.228969928.8719637428.8719637429.95199843627.0594646727.3913756728.9940979528.9940979530.02645979727.0859319927.5381769129.0668273929.0668273930.07906451827.1061101327.6522718629.1200739729.1200739730.12157731927.1228182627.7140045329.1613306629.1613306630.156393431027.1373769627.75053144

20、29.1940453329.1940453330.186802161127.1490560827.7775831929.2226321529.2226321530.210415411227.1577157427.7989257129.2453846429.2453846430.2300561327.1637220927.8156146729.2635058229.2635058230.246126521427.1678847827.8291036829.2792735729.2792735730.260603891527.1705728827.8416911929.2938893229.293

21、8893230.272715751627.1727715127.8526110829.3059162929.3059162930.283745141727.1741701727.8618996729.3169821229.3169821230.293176611827.1752856527.8695644829.3274565229.3274565230.300863661927.1764373627.8763970129.3370857329.3370857330.307920282027.1774610727.8829755829.3465949429.3465949430.3147873

22、8測試PSNR126.965327.584528.213128.839129.7407碼書大小對量化質(zhì)量的而影響由上圖可以得出的結(jié)論是碼書越大量化質(zhì)量越好，但是碼書不能夠太大，因?yàn)榇a書太大的話量化的時間會很長。圖像對比（最終碼書量化，再重建得到）：分離平均碼書：分離平均碼書（64）誤差次數(shù)Airplanebaboonlenapeppers訓(xùn)練MSE17170.7300634318.5805983379.9455384290.774532408.6160145915.0878605558.5858806794.43624643253.2134215693.5151913316.657750837

23、3.03316734219.4723797555.8604436264.1605335293.80451225192.9943612471.4405852231.9167301241.90826446173.135918422.4208913206.4811977206.78857947161.1384795390.511837185.1944224182.43904828152.2601668371.4330576168.1937008164.2822429143.9787012356.9006612152.334331148.683667610127.9232446345.95698381

24、39.6102858133.029061511114.1083504337.8509589131.6691921123.653861212107.8492675331.0789827126.2852785117.498296913104.1289049325.4472044121.5952912112.810304214101.1118301320.4654451117.5155167108.76747251598.33314639315.89324114.3662752105.32886811696.01235123311.6675322111.6469659102.34621711794.

258818488109.021111399.630460471892.5430899304.5051295106.418226996.9739370919912510133104.147923194.867837362089.90987003298.1735654102.064826293.30150953PSNR19.57516986811.7773933212.841711.80544667222.0176497718.5161756720.6599040719.1302131324.0959363919.7202438223.1249023

26、822.41332913424.7170048820.6811459123.9121242823.450219525.2753574121.3965339424.4774828224.29429655625.7469318721.8733497224.981998524.97553811726.05881122.2144615825.4545245825.51962563826.304940622.4319980725.8727063425.9748974926.5478210922.6053300826.3028257126.408170961027.0613089522.740582592

27、6.6816294526.891338341127.5576293422.8435520526.935961927.208726791227.8026316122.9314874927.1172763527.430487891327.955090623.0059981627.2816360427.607315911428.082783923.0729915327.4298514727.765813241528.2038042523.1354002927.5478238427.905329441628.3075325623.1938879927.6523343528.030085661728.3

28、915743323.2469627527.7556975628.146882231828.4673636523.2948574827.8606434228.264253331928.5318405923.3415184427.9542974728.359613612028.5927299123.3861122228.042042628.43191691測試PSNR125.405327.584524.664224.6027分離均值碼書與隨機(jī)碼書比較：可以看出，隨機(jī)生成的碼書在迭代次數(shù)較少的情況下明顯優(yōu)于分離平均得到的碼書，隨著迭代次數(shù)的增加，兩者的差別逐漸減小。圖像比較（最終碼書量化，在重構(gòu)得到

29、）：提高部分利用均值設(shè)計(jì)初始碼書方法（64）誤差次數(shù)Airplanebaboonlenapeppers訓(xùn)練MSE17171.4458014319.7077883384.6664484291.7272752527.7222753990.0796421673.9863813.49430023330.9812565798.8849011483.2866254424.57599384289.0571756692.4763432399.2372155357.25213765268.1722185618.8244011344.63625823276207402573.916775

30、6301.9553754295.59834617241.2950596546.9347016272.9361046264.9316158230.3776273531.0919474251.4228986241.23950129223.5256896520.0907886232.0566451225.124475810218.9752149510.9297034216.2635098213.276478211215.1240114503.2871885201.4270405203.950406312211.2820938495.9935345189.3415554195.08664313207.

31、7133647489.5776563182.0066314187.526860714204.145619484.3502194177.6835003181.992688615201.103494480.0944851174.2528687177.996263416198.4930522476.5926357171.2202945174.895363917195.9812684473.7328361168.561951172.436085218194.0032635471.2349858166.1058917170.113620219192.6098096469.062551163.877474168.139567820191.310611467.2206607161.778591166.1694625PSNR19.57473640311.7762599112.8356448511.80448245220.9067493518.17