鄧小煉，杜玉琪，王長(zhǎng)耀，王曉花

?

像素自相關(guān)矩陣的閾值自適應(yīng)角點(diǎn)檢測(cè)算法

鄧小煉1，杜玉琪1，王長(zhǎng)耀2，王曉花1

（1. 三峽大學(xué)理學(xué)院，宜昌 443002； 2. 中國(guó)科學(xué)院遙感與數(shù)字地球研究所，北京 100101）

針對(duì)Harris角點(diǎn)檢測(cè)算法中角點(diǎn)響應(yīng)函數(shù)（corner response function，CRF）系數(shù)閾值與非極大值抑制系數(shù)閾值需要人為設(shè)定所造成的可變性和隨機(jī)性等問(wèn)題，該文提出一種通過(guò)計(jì)算圖像每個(gè)像素的自相關(guān)矩陣行列式值，構(gòu)造特征角點(diǎn)圖像進(jìn)行自適應(yīng)閾值分割的改進(jìn)Harris角點(diǎn)檢測(cè)算法。該算法首先通過(guò)計(jì)算原圖像經(jīng)過(guò)方向?yàn)V波和低通濾波后各像素的自相關(guān)矩陣行列式值，以此構(gòu)造特征角點(diǎn)圖像；然后采用OTSU算法計(jì)算特征角點(diǎn)圖像分割閾值，從而篩選出預(yù)選區(qū)域；最后結(jié)合改進(jìn)的非極大值抑制方法提取有效角點(diǎn)。通過(guò)5組角點(diǎn)檢測(cè)對(duì)比試驗(yàn)結(jié)果數(shù)據(jù)分析，不同類型圖像的角點(diǎn)檢測(cè)準(zhǔn)確率均有提高，高分二號(hào)遙感影像的角點(diǎn)檢測(cè)準(zhǔn)確率提高27.06個(gè)百分點(diǎn)，可以初步得出，該算法相比傳統(tǒng)Harris角點(diǎn)檢測(cè)算法不但能夠自動(dòng)計(jì)算角點(diǎn)檢測(cè)的最佳閾值，而且能夠更準(zhǔn)確地定位角點(diǎn)和去除邊緣偽角點(diǎn)，從而提高了角點(diǎn)檢測(cè)的精確度，該研究可為農(nóng)業(yè)遙感影像數(shù)據(jù)檢測(cè)提供參考。

圖像處理；算法；角點(diǎn)檢測(cè)；自相關(guān)矩陣；特征角點(diǎn)圖像；非極大值抑制

0 引 言

角點(diǎn)通常是指圖像中梯度值與梯度變化率都非常大的像素點(diǎn)以及圖像邊緣曲線曲率極大的像素點(diǎn)，它能反映出圖像的局部特征，在降低信息數(shù)據(jù)量的同時(shí)有效地保留了圖像的重要特征[1]。角點(diǎn)檢測(cè)在目標(biāo)識(shí)別[2]與跟蹤、光流計(jì)算、圖像配準(zhǔn)[3]、相機(jī)標(biāo)定、運(yùn)動(dòng)估計(jì)、三維測(cè)量[4]與重構(gòu)等計(jì)算機(jī)視覺(jué)處理中發(fā)揮著極其重要的作用。此外，圖像配準(zhǔn)和鑲嵌的自動(dòng)化是農(nóng)作物長(zhǎng)勢(shì)遙感監(jiān)測(cè)和信息快速獲取的關(guān)鍵技術(shù)之一；遙感變化檢測(cè)技術(shù)也是耕地快速高效提取和動(dòng)態(tài)監(jiān)測(cè)方法的關(guān)鍵技術(shù)之一。因此，角點(diǎn)檢測(cè)在農(nóng)業(yè)信息化領(lǐng)域也發(fā)揮著極其關(guān)鍵作用。

目前，角點(diǎn)檢測(cè)的方法主要分為兩類：基于圖像灰度信息的方法和基于圖像邊緣輪廓特征[5]的方法。前者主要通過(guò)計(jì)算曲率[6]及梯度進(jìn)行角點(diǎn)檢測(cè)，如SUSAN算法[7]，Harris算法[8-9]和Morevec算法[10]。而后者需要對(duì)圖像預(yù)分割、輪廓鏈碼提取和角點(diǎn)檢測(cè)[11-12]3個(gè)步驟，如基于邊界鏈碼的角點(diǎn)檢測(cè)[12-14]，基于邊界曲率的角點(diǎn)檢測(cè)[15]，基于邊界小波變換的角點(diǎn)檢測(cè)[16]。在基于圖像灰度信息的方法中，Harris角點(diǎn)檢測(cè)算法效果較為理想，在紋理較豐富的區(qū)域可提取大量特征點(diǎn)，對(duì)圖像旋轉(zhuǎn)、灰度變化、噪音和視點(diǎn)變換不敏感。眾多學(xué)者在此基礎(chǔ)上進(jìn)行了改進(jìn)，如趙萌等[17]提出一種針對(duì)高斯函數(shù)參數(shù)值自適應(yīng)的Harris算法，但該算法需要在選定區(qū)域中設(shè)置多個(gè)值，并對(duì)不同值生成的預(yù)選角點(diǎn)響應(yīng)函數(shù)采用約束準(zhǔn)則篩選出最大值，算法較為復(fù)雜且運(yùn)算時(shí)間較長(zhǎng)；Mikolajczyk等[18]構(gòu)建了具有尺寸不變的Harris-Laplace算子和仿射不變的Harris-affine算子；王玉珠等[19]應(yīng)用B樣條函數(shù)替代了Harris算法中的高斯函數(shù)；Gevrekic等[20]提出的方法改進(jìn)了Harris算法在光照改變條件下的穩(wěn)定性；龍中杰等[21]應(yīng)用領(lǐng)域像素取差法和Susan去角點(diǎn)的特性改進(jìn)Harris算法；毛雁明等[22]提出雙閾值法解決了非極大值抑制時(shí)不易設(shè)置閾值的問(wèn)題；李海等[23]使用直線檢測(cè)算法來(lái)進(jìn)行角點(diǎn)的自動(dòng)提取，在進(jìn)行非極大值抑制時(shí)回避了閾值設(shè)定；Mokhtarian等[24]提出CSS(curvature scale space)算法圖像進(jìn)行幾何變換時(shí)相當(dāng)穩(wěn)定，在多尺度空間下也能具有很好的魯棒性；Awrangjeb等[25]提出了基于CPDA(chord-point distance accumulation)角點(diǎn)檢測(cè)算法，采用點(diǎn)到弦距離累加CPDA來(lái)代替曲率的計(jì)算；周志艷等[26]提出的Harris角點(diǎn)檢測(cè)算法針對(duì)角點(diǎn)響應(yīng)函數(shù)進(jìn)行改進(jìn)，然而圖像灰度標(biāo)準(zhǔn)差的選擇對(duì)角點(diǎn)的數(shù)量以及角點(diǎn)的匹配精度具有很大的影響。

本文針對(duì)Harris算法中角點(diǎn)響應(yīng)函數(shù)系數(shù)閾值和非極大值抑制系數(shù)閾值的設(shè)定具有很大的隨機(jī)性和動(dòng)態(tài)可變性，容易造成角點(diǎn)檢測(cè)效果不理想，因此提出一種新的思路：直接利用圖像像素的自相關(guān)矩陣計(jì)算得出特征角點(diǎn)圖像，之后進(jìn)行自適應(yīng)的閾值分割，并采用改進(jìn)的非極大值抑制方法對(duì)預(yù)篩選結(jié)果進(jìn)行處理，最終檢測(cè)出具有較高定位精度的最佳角點(diǎn)。此外，除了常見(jiàn)的JPG、BMP等圖像格式之外，本文改進(jìn)算法更適用于遙感多光譜影像，因?yàn)檫b感影像的波段數(shù)較多，圖像特征相比通用圖像格式更多，可以反映出更多的在可見(jiàn)光波段無(wú)法反映的地物光譜特征，因此更有利于角點(diǎn)檢測(cè)。

1 Harris角點(diǎn)檢測(cè)原理

Harris角點(diǎn)檢測(cè)算法是通過(guò)微分運(yùn)算和自相關(guān)矩陣來(lái)檢測(cè)角點(diǎn)的，其中微分運(yùn)算能夠反映像素點(diǎn)在任意方向灰度的變化量[27]，因而能夠有效地區(qū)分出角點(diǎn)和邊緣點(diǎn)。設(shè)像素點(diǎn)(,)的灰度值為(,)。若以像素點(diǎn)(,)為中心的小窗口，在,方向上分別移動(dòng)微小的位移,，則灰度(,)變化的表達(dá)式為

()的矩陣形式為

由以上各式得出(,)的性質(zhì)主要由自相關(guān)矩陣決定，由于自相關(guān)矩陣的特征值不易計(jì)算，通常選擇計(jì)算每個(gè)像素點(diǎn)的角點(diǎn)響應(yīng)函數(shù)CRF

系數(shù)為Harris算法的經(jīng)驗(yàn)常數(shù)，一般為0.04～0.06。以每個(gè)像素點(diǎn)為中心取其3×3的鄰域，若中心像素點(diǎn)的角點(diǎn)響應(yīng)函數(shù)值大于某個(gè)設(shè)定的閾值并且是該鄰域的最大值，則該像素點(diǎn)可判定為角點(diǎn)[28]。

2 改進(jìn)的自適應(yīng)Harris角點(diǎn)檢測(cè)算法

2.1 改進(jìn)算法的思想

Harris角點(diǎn)檢測(cè)算法是一種經(jīng)典的角點(diǎn)檢測(cè)算法，雖然具有旋轉(zhuǎn)和仿射變換不變性，但在實(shí)際應(yīng)用中仍存在一些不足：1）Harris算法在提取角點(diǎn)時(shí)，需要在設(shè)定閾值的基礎(chǔ)上進(jìn)行非極大值抑制，而且閾值的設(shè)置決定了角點(diǎn)提取的準(zhǔn)確度，閾值設(shè)定偏大則會(huì)引起角點(diǎn)的丟失，偏小則會(huì)導(dǎo)致提取大量偽角點(diǎn)；2）Harris算法中的角點(diǎn)響應(yīng)函數(shù)中的系數(shù)需要隨圖像的變化進(jìn)行人為調(diào)整，智能化程度低，使其在實(shí)際應(yīng)用中效率降低。針對(duì)上述問(wèn)題，本文提出一種基于圖像像素自相關(guān)矩陣的改進(jìn)的Harris角點(diǎn)檢測(cè)算法：首先，對(duì)圖像做方向?yàn)V波，再進(jìn)行高斯平滑，減少噪聲點(diǎn)的干擾；然后計(jì)算圖像像素的自相關(guān)矩陣行列式，以此構(gòu)造特征角點(diǎn)圖像；最后對(duì)特征角點(diǎn)圖像進(jìn)行OTSU閾值分割篩選出預(yù)選區(qū)域，結(jié)合改進(jìn)的非極大值抑制方法提取鄰域最佳角點(diǎn)。該算法舍棄了經(jīng)典Harris算法中的角點(diǎn)響應(yīng)函數(shù)和非極大值抑制閾值2個(gè)參數(shù)，直接從與圖像角點(diǎn)檢測(cè)真正相關(guān)的像素自相關(guān)矩陣入手，避免了Harris算法需要設(shè)定2個(gè)閾值的隨機(jī)性和不確定性等缺點(diǎn)。無(wú)需人為調(diào)整閾值大小，通過(guò)OTSU計(jì)算得出自適應(yīng)閾值，以期提高工作效率和角點(diǎn)檢測(cè)精度，突出體現(xiàn)角點(diǎn)檢測(cè)智能化的特點(diǎn)。具體改進(jìn)算法的流程圖如圖1所示。

圖1 基于像素自相關(guān)矩陣的角點(diǎn)檢測(cè)算法流程圖

2.2 改進(jìn)算法的實(shí)現(xiàn)步驟

由于Harris角點(diǎn)檢測(cè)響應(yīng)函數(shù)CRF受系數(shù)影響，所以角點(diǎn)提取的效果依賴于系數(shù)的設(shè)定，因此對(duì)于某一幅圖像可能需要多次人工交互調(diào)整系數(shù)才能獲得理想的檢測(cè)結(jié)果?？紤]到人為設(shè)定系數(shù)的隨機(jī)性，本文對(duì)經(jīng)典的Harris角點(diǎn)檢測(cè)算法進(jìn)行了改進(jìn)，通過(guò)計(jì)算圖像像素的自相關(guān)矩陣行列式來(lái)構(gòu)造特征角點(diǎn)圖像，在此基礎(chǔ)上通過(guò)OTSU算法計(jì)算特征角點(diǎn)圖像的自適應(yīng)分割閾值來(lái)篩選出角點(diǎn)預(yù)選區(qū)域，最后采用改進(jìn)的非極大值抑制方法來(lái)判定鄰域最佳角點(diǎn)，這樣避免了人為設(shè)定所造成的誤差，提高了檢測(cè)效率與精確度。

2.2.1 計(jì)算像素自相關(guān)矩陣構(gòu)造特征角點(diǎn)圖像

計(jì)算灰度圖像上像素點(diǎn)(,)在水平方向與垂直方向上的梯度、及其乘積，通過(guò)高斯平滑函數(shù)(,)對(duì)其進(jìn)行濾波計(jì)算，得出圖像像素的自相關(guān)矩陣，在此應(yīng)該注意掌握高斯平滑窗口的尺寸，以不超過(guò)7×7為宜。若窗口選擇偏大在卷積運(yùn)算過(guò)程中會(huì)導(dǎo)致角點(diǎn)位置的偏移；若窗口選擇偏小會(huì)由于噪聲點(diǎn)的影響呈現(xiàn)出較多的偽角點(diǎn)。計(jì)算圖像每個(gè)像素自相關(guān)矩陣的行列式，在此基礎(chǔ)上構(gòu)造特征角點(diǎn)圖像。

2.2.2 對(duì)特征角點(diǎn)圖像進(jìn)行OTSU閾值分割

1類和2類以及整幅圖像的灰度均值1，2，T分別為

1類和2類的方差分別為

以此對(duì)特征角點(diǎn)圖像進(jìn)行閾值分割，篩選出預(yù)選區(qū)域。

2.2.3 使用非極大值抑制方法提取最佳角點(diǎn)

為了提高檢測(cè)效率以及準(zhǔn)確性，本文根據(jù)Neubeck[30]等改變模板窗口大小以及分塊處理的方法，采用改進(jìn)的非極大值抑制算法對(duì)圖像預(yù)選區(qū)域進(jìn)行閾值分割并提取角點(diǎn)，具體算法步驟如下。

首先，計(jì)算圖像像素自相關(guān)矩陣行列式得出圖像的梯度變化劇烈區(qū)域，將這些預(yù)選區(qū)域劃分成多個(gè)尺寸為(+1)×(+1) 像素的正方形子區(qū)，如圖2所示；其次，設(shè)定一個(gè)窗口為3×3像素的正方形模板，在該子區(qū)中對(duì)應(yīng)的每個(gè)正方形模板搜索梯度為極大值的候選點(diǎn)，圖2中的黑點(diǎn)表示該子區(qū)中篩選出的候選點(diǎn)集；最后，對(duì)該子區(qū)中的候選點(diǎn)集結(jié)合自適應(yīng)閾值做完整鄰域的角點(diǎn)檢測(cè)。依照上述方法對(duì)整體預(yù)選區(qū)域執(zhí)行非極大值抑制，剔除邊緣上的偽角點(diǎn)，從中檢測(cè)出圖像的最佳角點(diǎn)。該方法可以減少運(yùn)算量，快速精準(zhǔn)地檢測(cè)出角點(diǎn)。

圖2 改進(jìn)的非極大值抑制方法示意圖

3 試驗(yàn)與結(jié)果分析

3.1 試驗(yàn)方法

為了驗(yàn)證本算法角點(diǎn)檢測(cè)的性能，選擇角點(diǎn)特征比較明顯的5幅圖像，包括經(jīng)典的立方體圖像，某高校宿舍樓圖像，某大廈圖像，以及高分二號(hào)遙感圖像作為數(shù)據(jù)源，采用經(jīng)典的Harris算法和本文提出的改進(jìn)算法進(jìn)行仿真試驗(yàn)，并對(duì)檢測(cè)結(jié)果進(jìn)行數(shù)據(jù)對(duì)比分析。試驗(yàn)環(huán)境為Inter(R) Core(TM)2處理器、2GB內(nèi)存、64位Window7操作系統(tǒng)、Matlab R2014b

3.2 試驗(yàn)結(jié)果

角點(diǎn)檢測(cè)試驗(yàn)結(jié)果如圖3～圖7所示。圖3－圖7分別為針對(duì)立方體、高校宿舍樓、建筑大廈圖像及高分二號(hào)遙感影像進(jìn)行的角點(diǎn)提取的對(duì)比試驗(yàn)結(jié)果。

圖3 立方體圖像角點(diǎn)檢測(cè)算法結(jié)果對(duì)比

圖4是采用高校宿舍樓圖像進(jìn)行角點(diǎn)提取的對(duì)比試驗(yàn)結(jié)果，其中圖4a為原圖，圖4b為對(duì)原圖進(jìn)行自相關(guān)行列式計(jì)算得出的特征角點(diǎn)圖像，圖4c為采用經(jīng)典的Harris算法和本文改進(jìn)算法提取的角點(diǎn)差異圖，圖4d為局部差異圖。圖4c和圖4d中紅色十字為Harris算法提取結(jié)果，綠色圓圈為本文算法提取結(jié)果。

注：圖c, d中紅色十字為Harris算法提取結(jié)果，綠色圓圈為本文算法提取結(jié)果。下同。

圖5 Harris與本文算法高層建筑物圖像角點(diǎn)檢測(cè)結(jié)果對(duì)比

除了通用的圖像格式數(shù)據(jù)外，針對(duì)角點(diǎn)檢測(cè)是遙感圖像配準(zhǔn)和鑲嵌自動(dòng)化的關(guān)鍵步驟，本文進(jìn)行了針對(duì)高分二號(hào)遙感影像的2組試驗(yàn)（圖6，圖7）。其中圖6為高分二號(hào)遙感影像GF2_PMS2_E116.7_N40.0_20160609_L1A0001633143北京首都國(guó)際機(jī)場(chǎng)的部分圖像開(kāi)展角點(diǎn)檢測(cè)試驗(yàn)。圖7是應(yīng)用高分二號(hào)遙感影像GF2_PMS1_E82.7_N44.7_20160626_L1A0001666662中農(nóng)作物植被的部分影像進(jìn)行的角點(diǎn)提取的對(duì)比試驗(yàn)結(jié)果。

圖6 Harris與本文算法高分二號(hào)遙感融合影像角點(diǎn)檢測(cè)結(jié)果對(duì)比

圖7 農(nóng)作物植被部分遙感影像Harris算法與本文算法角點(diǎn)檢測(cè)結(jié)果對(duì)比

從以上2種算法的角點(diǎn)檢測(cè)結(jié)果差異圖中可以直觀看出本算法角點(diǎn)檢測(cè)結(jié)果優(yōu)于經(jīng)典Harris算法，具體數(shù)據(jù)見(jiàn)表1和表2列出。其中表1包含圖像不同算法的重要參數(shù)，表2中的標(biāo)志點(diǎn)數(shù)表示標(biāo)記出的角點(diǎn)數(shù)，正確數(shù)表示檢測(cè)出的真正角點(diǎn)數(shù)，漏檢數(shù)表示未被檢測(cè)出的角點(diǎn)數(shù)，誤檢數(shù)表示檢測(cè)出的偽角點(diǎn)數(shù)，準(zhǔn)確率為檢測(cè)出的角點(diǎn)數(shù)與圖像總角點(diǎn)數(shù)之比。

表1 所有附圖參數(shù)一覽表

表2 角點(diǎn)檢測(cè)統(tǒng)計(jì)

從以上對(duì)比試驗(yàn)結(jié)果以及表格分析可得出，圖3為簡(jiǎn)易的立體幾何圖形，其Harris算法的準(zhǔn)確率比改進(jìn)算法低30.51個(gè)百分點(diǎn)，圖4和圖5為建筑物圖像，其Harris算法的準(zhǔn)確率比改進(jìn)算法分別低13.36個(gè)百分點(diǎn)和3.07個(gè)百分點(diǎn)，圖6和圖7為高分二號(hào)遙感影像，其Harris算法的準(zhǔn)確率比改進(jìn)算法分別低7.55個(gè)百分點(diǎn)和27.06個(gè)百分點(diǎn)，對(duì)于不同類型的圖像，改進(jìn)算法的準(zhǔn)確率均有提升。Harris算法容易誤提取出邊緣上的偽角點(diǎn)，角點(diǎn)提取的準(zhǔn)確率不高，并且角點(diǎn)定位存在偏差。這些缺點(diǎn)取決于Harris算法采用了較為單一固定的系數(shù)和非極大值抑制系數(shù)的閾值，使該算法只能針對(duì)某些特定圖像提取較好的結(jié)果，而在實(shí)際應(yīng)用中，并不適應(yīng)其他圖像。若要改善提取效果，則需反復(fù)試驗(yàn)更改2個(gè)閾值，增大工作量降低運(yùn)算效率。本算法避免了閾值的反復(fù)設(shè)定，應(yīng)用OTSU算法直接從特征角點(diǎn)圖像中計(jì)算出分割閾值，從而篩選出預(yù)選區(qū)域，最后通過(guò)改進(jìn)的非極大值抑制方法檢測(cè)出最佳角點(diǎn)。改進(jìn)的算法具有較好的智能性，避免了設(shè)定Harris算法中角點(diǎn)響應(yīng)函數(shù)系數(shù)閾值與大抑制系數(shù)閾值的隨機(jī)性和主觀性，能夠獲得更多的有效角點(diǎn)，減少了偽角點(diǎn)的提取，并且具有較高的定位精度。從圖4d、圖5d、圖6d中可以明顯看出，本論文算法較傳統(tǒng)算法能夠更準(zhǔn)確地定位角點(diǎn)。圖3d中的六棱柱最上面的2個(gè)角點(diǎn)均未能檢出，主要原因是圖像已經(jīng)過(guò)拉伸處理，原圖中這兩個(gè)角點(diǎn)的灰度值很低，在單閾值情況下較難檢測(cè)出來(lái)，可以考慮分窗口檢測(cè)，每個(gè)窗口單獨(dú)計(jì)算閾值，這樣提取效果會(huì)更好些。圖5c和圖6c的Harris算法角點(diǎn)提取總數(shù)雖然比本文算法的角點(diǎn)提取總數(shù)多，但是很多角點(diǎn)都是位于圖像的邊緣輪廓，并非圖像中有價(jià)值的角點(diǎn)，相比而言，本文算法所提取出的角點(diǎn)更有實(shí)際價(jià)值。

4 結(jié) 論

本文針對(duì)經(jīng)典Harris角點(diǎn)檢測(cè)算法中角點(diǎn)響應(yīng)函數(shù)的系數(shù)和閾值的隨機(jī)性以及需要設(shè)定非極大值抑制系數(shù)閾值等方面的不足，提出采用自相關(guān)矩陣和自適應(yīng)閾值相結(jié)合的改進(jìn)算法。應(yīng)用自相關(guān)矩陣對(duì)圖像進(jìn)行處理，提取特征角點(diǎn)圖像，在此基礎(chǔ)上應(yīng)用OTSU分割算法得到自適應(yīng)閾值，避免了人為設(shè)定閾值的隨機(jī)性和反復(fù)性，并提出改進(jìn)的非極大值抑制方法。與經(jīng)典的Harris算法相比較，本算法有如下幾個(gè)優(yōu)勢(shì)：1）通過(guò)直接計(jì)算圖像像素的自相關(guān)矩陣行列式來(lái)構(gòu)造特征角點(diǎn)圖像，改進(jìn)優(yōu)化了角點(diǎn)檢測(cè)模型；2）可以不依賴于人為設(shè)定參數(shù)，提高了角點(diǎn)檢測(cè)的智能性和時(shí)效性；3）能夠較準(zhǔn)確地定位角點(diǎn)并剔除偽角點(diǎn)，具有更好的檢測(cè)性能，針對(duì)農(nóng)作物植被遙感圖像角點(diǎn)檢測(cè)的準(zhǔn)確率校經(jīng)典Harris算法提高27.06個(gè)百分點(diǎn)；4）能夠針對(duì)多種圖像格式進(jìn)行角點(diǎn)檢測(cè)，具有更強(qiáng)的適應(yīng)性。如何更為客觀地評(píng)價(jià)角點(diǎn)檢測(cè)結(jié)果，構(gòu)造更為有效的角點(diǎn)檢測(cè)分析模型，拓展其在農(nóng)業(yè)信息化領(lǐng)域的作用，也是值得進(jìn)一步研究的方向。

[1] Chen Jie, Zou Lihui, Zhang Juan, et al. The comparison and application of corner detection algorithms[J]. Journal of Multimedia, 2009, 4(6): 435－441.

[2] Paul J, Oechslein B, Erhardt C, et al. Self-adaptive corner detection on MPSoC through resource-aware programming original[J]. Journal of Systems Architecture the Euromicro Journal, 2015, 61(10): 520－530.

[3] 姚國(guó)標(biāo)，鄧喀中，張力，等. 基于 Harris-Affine 的寬基線立體影像LSM匹配方法[J]. 中南大學(xué)學(xué)報(bào)，2014，45(8)：2661－2668.

Yao Guobiao, Deng Kazhong, Zhang Li, et al. Least square matching method for wide baseline stereo images based on Harris-Affine features[J]. Journal of Central South University, 2014, 45(8): 2661－2668. (in Chinese with English abstract)

[4] Mokhtarian F, Mohanna F. Performance evaluation of corner detectors using consistency and accuracy measures[J]. Computer Vision and Image Understanding, 2006, 102(1): 81－94.

[5] 邢遠(yuǎn)秀，章登義，趙儉輝. 利用多尺度弦角尖銳度累積的自適應(yīng)角點(diǎn)檢測(cè)算子[J]. 武漢大學(xué)學(xué)報(bào)信息科學(xué)版，2015，40(5)：617－627.

Xing Yuanxiu, Zhang Dengyi, Zhao Jianhui. An adaptive threshold corner detector based on multi-scale chord-angle sharpness accumulation[J]. Geomatics and Information Science of Wuhan University, 2015, 40(5): 617－627. (in Chinese with English abstract)

[6] Chen Suting, Meng Hao, Zhang Chuang, et al. A KD curvature based corner detector[J]. Neurocomputing, 2016, 172(2): 434－441.

[7] Smith S M, Bpady J M. SUSAN-A new approach to low level image processing[J]. International Journal of Computer Vision, 1997, 23(1): 45－47.

[8] Harris C G, Stephes M J. A combined corner and edge detector[C]//Proceeding of Fourth Alvey Vision Conference, 1988: 147－151.

[9] Cordelia Schmid, Rrger Mohr. Christian bauckhage evaluation of interest point detectors[J]. International Journal of Computer Vision, 2000, 37(2): 151－172.

[10] Moravec H P. Towards automatic visual obstacle avoidance[C]//Proceedings of the 5th International Joint Conference on Artificial Intelligence, 1977: 584－590.

[11] 趙斌，周軍. 基于改進(jìn)棋盤的角點(diǎn)自動(dòng)檢測(cè)與排序[J]. 光學(xué)精密工程，2015，23(1)：237－244.

Zhao Bin, Zhou Jun. Automatic detection and sorting of corners by improved chessboard pattern[J]. Opt. precision Eng, 2015, 23(1): 237－244. (in Chinese with English abstract)

[12] Zhang W C, Shui P L. Contour-based corner detection via angle difference of principal directions of anisotropic Gaussian directional derivatives[J]. Pattern Recognition, 2015, 48(9): 2785－2797.

[13] 陳華偉，吳祿慎，袁小翠. 基于鏈碼改進(jìn)算法的鋼軌表面缺陷識(shí)別[J]. 計(jì)算機(jī)工程與設(shè)計(jì)，2015，36(11)：3097－3101. Chen Haiwei, Wu Lushen, Yuan Xiaocui. Rail surface defect detection based on improved chain-code algorithm[J]. Computer Engineering and Design, 2015, 36(11): 3097－3101. (in Chinese with English abstract)

[14] 李振華，江耿紅，徐勝男，等. 基于輪廓多邊形擬合的紅外與可見(jiàn)光圖像配準(zhǔn)算法[J]. 系統(tǒng)工程與電子技術(shù)，2015，37(12)：2872－2878.

Li Zhenhua, Jiang Genghong, Xu Shengnan, et al. Image registration algorithm for infrared and visible images based on contour polygon fitting[J]. Systems Engineering and Electronics, 2015, 37(12): 2872－2878. (in Chinese with English abstract)

[15] Mokhtarian F, Bober M. Robust image corner detection through curvature scale space[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1988, 6(1): 121－128.

[16] 侯北平，李平，宋執(zhí)環(huán). 紙漿纖維的形狀特征提取應(yīng)用研究[J]. 浙江大學(xué)學(xué)報(bào)：工學(xué)版，2006，40(7)：1132－1136.

Hou Beiping, Li Ping, Song Zhihuan. Applied study on paper fibre shape characteristics extraction[J]. Journal of Zhejiang University: Engineering Science, 2006, 40(7): 1132－1136. (in Chinese with English abstract)

[17] 趙萌，溫佩芝，鄧星，等. 一種參數(shù)自適應(yīng)的Harris角點(diǎn)檢測(cè)算法[J]. 桂林電子科技大學(xué)學(xué)報(bào)，2016，36(3)：215－219. Zhao Meng, Wen Peizhi, Deng Xing, et al. A parameter adaptive Harris corner detection algorithm[J]. Journal of Guilin University of Electronic Technology, 2016, 36(3): 215－219. (in Chinese with English abstract)

[18] Mikolajczyk K, Schmid C. Scale and affine invariant interest point detector[J]. International Journal of Computer Vision, 2004, 60(1): 63－86.

[19] 王玉珠，楊丹，張小洪. 基于B樣條的改進(jìn)型Harris角點(diǎn)檢測(cè)算法[J]. 計(jì)算機(jī)應(yīng)用研究，2007(2)：192－194.

Wang Yuzhu, Yang Dan, Zhang Xiaohong. Improved harris corner detection algorithm based on B-spline[J]. Application Research of Computers, 2007(2): 192－194. (in Chinese with English abstract)

[20] Gevrekci M, Gunturk B K. Reliable interest point detection under large illumination variations[C]//IEEE International conference on Image Processing, ICIP, 2008: 869－872.

[21] 龍中杰，王吉芳，左云波. 一種改進(jìn)的Harris與Susan相結(jié)合的角點(diǎn)檢測(cè)方法[J]. 計(jì)算機(jī)應(yīng)用與軟件，2013，30(12)：133－136.

Long Zhongjie, Wang Jifang, Zuo Yunbo. An improved corner detection algorithm with the combination of harris and susan[J]. Computer Applications and Software, 2013, 30(12): 133－136. (in Chinese with English abstract)

[22] 毛雁明，蘭美輝，王運(yùn)瓊，等. 一種改進(jìn)的基于Harris的角點(diǎn)檢測(cè)方法[J]. 計(jì)算機(jī)技術(shù)與發(fā)展，2009，19(5)：130－133.

Mao Yanming, Lan Meihui, Wang Yunqiong,et al. An improved corner detection method based on harris[J]. Computer Technology and Development, 2009, 19(5): 130－133. (in Chinese with English abstract)

[23] 李海，張憲民，陳忠. 基于直線檢測(cè)的棋盤格角點(diǎn)自動(dòng)提取[J]. 光學(xué)精密工程，2015，23(12)：3480－3489.

Li Hai, Zhang Xianming, Chen Zhong. Automatic corner detection of checkerboard based on LSD[J]. Opt.precision Eng2015, 23(12): 3480－3489. (in Chinese with English abstract)

[24] Mokhtarian F, Suomela R. Enhancing the curvature scale space corner detector[C]//Proc. of Scandinavian Conf. Image Analysis. Bergen, Norway: Int. Assoc. Pattern Recognition, 2001: 145－152.

[25] Awrangjeb M, Lu G J. Robust image corner detection based on the chord-to-point distance accumulation technique[J]. IEEE Transactions on Multimedia, 2008, 10(6): 1059－1072.

[26] 周志艷，閆夢(mèng)璐，陳盛德，等. Harris 角點(diǎn)自適應(yīng)檢測(cè)的水稻低空遙感圖像配準(zhǔn)與拼接算法[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2015，31(14)：186－193.

Zhou Zhiyan, Yan Menglu, Chen Shengde, et al. Image registration and stitching algorithm of rice low-altitude remote sensing based on Harris corner self-adaptive detection[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(14): 186－193. (in Chinese with English abstract)

[27] 鄧淇元，曲長(zhǎng)文，江源. 基于圓環(huán)模板的改進(jìn)Harris角點(diǎn)檢測(cè)算法[J]. 系統(tǒng)工程與電子技術(shù)，2016，38(4)：949－954. Deng Qiyuan, Qu Changwen, Jiang Yuan. Improved corner detection algorithm based on circle mask via harris[J]. Systems Engineering and Electronics, 2016, 38(4): 949－954. (in Chinese with English abstract)

[28] 吳鵬，徐洪玲，李雯霖，等. 基于區(qū)域檢測(cè)的多尺度Harris角點(diǎn)檢測(cè)算法點(diǎn)檢測(cè)[J]. 哈爾濱工程大學(xué)學(xué)報(bào)，2016，37(7)：969－973.

Wu Peng, Xu Hongling, Li Wenlin, et al. Multi-scale Harris- corner detection algorithm based on region detection[J]. Journal of Harbin Engineering University, 2016, 37(7): 969－973. (in Chinese with English abstract)

[29] Otsu N. A threshold selection method from gray-level histograms[J]. IEEE Transactions on Systems, Man and Cybernetics, 1979, 9(1): 62－66.

[30] Neubeck A, van Gool L. Efficient non-maximum suppression[C]//Proceedings of the 18th International Conference on Pattern Recognition, 2006, 3: 850－855.

An adaptive threshold corner detection algorithm based on auto-correlation matrix of image pixel

Deng Xiaolian1, Du Yuqi1, Wang Changyao2, Wang Xiaohua1

(1443002,;2100101,)

Harris algorithm is a classical corner detection algorithm. It can extract corners of image quickly and has a certain degree of anti-noise ability, but it has corner location error to some extent. It needs to artificially set 2 threshold parameters, and it can not easily eliminate false corners such as edge points, so it has somewhat lower accuracy of corner detection. For above-mentioned reasons, a modified Harris corner detection algorithm based on auto-correlation matrix of image pixel was proposed in this paper, and the purpose was not only to solve the problem of the variability and randomness of setting thresholds for corner response function (CRF) and non-maximum suppression in Harris algorithm, but also to improve the accuracy of corner location. In our paper, the most important innovation is embodied in 2 aspects: One is avoiding to set 2 thresholds of traditional Harris corner detection algorithm artificially, the other is locating corner more accurately by modified non-maximum suppression method. Firstly, original image was filtered by directional filtering and Gaussian low-pass filtering, and feature corner image was constructed by calculating determinant of every pixel’s auto-correlation matrix. Potential corners of image could be heightened effectively, which had more significant intensity than other surrounding pixels, and could be recognized easily in feature corner image. Secondly, in order to improve intelligent level of the modified algorithm, we selected adaptive OTSU algorithm to determine segmentation threshold. The segmentation threshold of feature corner image could be calculated by OTSU algorithm, and the pre-selected regions were obtained. So the search range of corner detection was significantly decreased. On the basis, an optimized non-maximum suppression method was adopted in our research, which could divide each pre-selected region into several 3×3 square subranges, and correct corners were extracted from potential corners of each square subrange, false corners were eliminated effectively. Finally, in order to validate the efficiency and reliability of the modified algorithm, 5 groups of comparison experiments were performed in our research. Five images, including generic image format (jpg, bmp), and multi-band remote sensing image format (GF-2 data), were selected to test performance of the modified algorithm and Harris algorithm, which contained the total of detection corners, the number of correct corners, the number of false corners, the number of omissive corners, and the detection rate of correct corners. According 5 groups of comparison experiments, the accuracy of corner detection in different types of images is improved, for crop vegetation remote sensing image, the accuracy of corner detection is improved by 27.06 percentage points. We can draw a conclusion that the improved algorithm can not only calculate the optimal threshold automatically, but also locate the corners more accurately. Therefore, our modified algorithm can greatly improve the precision of corner detection. The proposed algorithm is more accurate and efficient than traditional algorithm, its adaptive characteristic makes it easy to be integrated in an image processing system or image registration module, and it has higher feasibility and application value. Experiments also show that there is some insufficiency to be improved in our research, for example, some corners in picture of cubes could not be detected correctly with either our modified algorithm or Harris algorithm. In our future research, we propose to partition an image into several sub-image blocks, and independently determine each sub-image block’s segmentation threshold by OTSU algorithm, so that the corners not prominent in full image can be significantly strengthened in sub-image blocks, and can be detected correctly. The research could provide reference for agricultural remote sensing image data detection.

image processing; algorithms; corner detection; auto-correlation matrix; feature corner image; non-maximum suppression

10.11975/j.issn.1002-6819.2017.18.018

TP391.41

A

1002-6819(2017)-18-0134-07

2017-03-09

2017-08-11

遙感科學(xué)國(guó)家重點(diǎn)實(shí)驗(yàn)室課題(Y6Y00200KZ)

鄧小煉，男，重慶人，博士，副教授，主要從事遙感信息處理，模式識(shí)別，變化檢測(cè)等方面的研究。Email：345937408@qq.com