外文翻譯--基于dsp的通過(guò)局部特征實(shí)時(shí)物體識(shí)別嵌入式系統(tǒng)

1、<p><b>　　附錄A</b></p><p>　　Real-time object recognition using local features on a DSP-based embedded system</p><p><b>　　Abstract</b></p><p>　　In the last

2、 few years, object recognition has become one of the most popular tasks in computer vision. In particular, this was driven by the development of new powerful algorithms for local appearance based object recognition. So-c

3、alled ‘‘smart cameras’’ with enough power for decentralized image processing became more and more popular for all kinds of tasks, especially in the field of surveillance. Recognition is a very important tool as the robus

4、t recognition of suspicious vehicles, persons or </p><p>　　Keywords DSP ; Object recognition; Local features; Vocabulary tree</p><p>　　Introduction</p><p>　　Object recognition is o

5、ne of the most popular tasks in the field of computer vision. In the past decade, big efforts were made to build robust object recognition systems based on appearance features with local extent. For such a framework to b

6、e applicable in the real world several attributes are very important: insensitivity against rotation, illumination or view point changes, as well as real-time behavior and large-scale operation. Current systems already h

7、ave a lot of these properties and, th</p><p>　　In turn, recently embedded vision platforms such as smart cameras have successfully emerged, however, only offering a limited amount of computational and memory

8、 resources. Nevertheless, embedded vision systems are already present in our everyday life. Almost everyone’s mobile phone is equipped with a camera and, thus, can be treated as a small embedded vision system. Clearly th

9、is gives rise to new applications, like navigation tools for visually impaired persons, or collaborative public monitori</p><p>　　For the reasons already mentioned, recognition tasks are a very important are

10、a of research. However, in this respect some attributes of embedded platforms strictly limit the practicability of current state-of-the-art approaches. For example, the amount of memory available on a device strictly lim

11、its the number of objects in the database. Therefore, for building an embedded object recognition system, one goal is to make the amount of data to represent a single object as small as possible in order</p><p

12、>　　In this work, we describe a method to deploy a medium sized object recognition system on a prototypical DSP based embedded platform. To the best of our knowledge, we are the first to extensively investigate issues

13、related to object recognition in the context of Embedded Systems; by now this is the only work studying the influence of various parameters on recognition performance and runtime behavior. We pick a set of high-level alg

14、orithms to describe objects by a set of appearance features. As a </p><p>　　The remainder of this paper is structured as follows. In Sect. 2 we give an overview about developments in both areas that we are b

15、ringing together in our work. On the one hand we list a number of references in the context of object recognition by computer vision; on the other hand, we cite a number of publications from the area of embedded smart se

16、nsors. A detailed description of the methods involved in building our object recognition algorithm is given in part 3. In Sect. 4 we outline our fram</p><p>　　Related Work</p><p>　　In the follow

17、ing we will give a short introduction to the topic of local feature based object recognition. Due to the huge amount of literature available, we will focus on the most promising approaches using local features, and refer

18、 to those algorithms which are somehow related to our work. We will also give a short overview about object recognition in the context of embedded systems, which, due to the sparseness of existing approaches, contain bot

19、h global and local methods, as well as algorith</p><p>　　Local-appearance based visual object recognition became popular after the development of powerful interest region detectors and descriptors. Early ful

20、l-featured object recognition systems dealing with all the individual algorithmic steps and their related problems were proposed by Schmid and Mohr, and Schiele and Crowley . The main idea behind local feature based obje

21、ct recognition is maintaining object representations from collections of locally sampled descriptions. In other words, the appeara</p><p>　　The collectivity of all descriptors from multiple objects(i.e., bag

22、s of descriptors) is used to build a database. Given this database and a new representation of an object to be recognized, correspondences are counted into a voting scheme to determine the correct match. Determining thes

23、e correspondences is a complex task. Descriptors are high dimensional feature vectors and matching a query descriptor means determining the exact nearest neighbors in the database. Unfortunately, by now, no algori</p&

24、gt;<p>　　The basic principle of interest points and regions is the search for spots and areas in an image which exhibit a predefined property making them special in relation to their local neighborhood. This prope

25、rty should make the region distinguishable from its neighborhood and detectable repeatedly. Furthermore, the detection of these features should be—to the best possible—illumination and viewpoint invariant.</p><

26、;p>　　The first important interest point detector, the so-called Harris Corner detector, was proposed in 1988 by Harris and Stephens. It exhibits excellent repeatability and was subsequently used for object recognition

27、 purposes by Schmid and Mohr. An extension to the Harris detector to include scale information was later reported by Mikolajczyk and Schmid as Harris–Laplace detector and was used by Schaffalitzky and Zisserman formulti-

28、view matching of unordered image sets. Another approach to detect bl</p><p>　　The currently most popular two-part approach known as scale invariant feature transform (SIFT) was proposed by Lowe, where the fi

29、rst part is an interest point detector. The DoG detector takes the differences of Gaussian blurred images as an approximation of the scale normalized Laplacian and uses the local maximum of the responses in scale space a

30、s an indicator for a keypoint. A complementary feature detector, the maximally stable extremal regions (MSER) detector, was proposed by Matas et al. In</p><p>　　Two affine covariant region detectors were pro

31、posed by Tuytelaars and Van Gool, intensity-based regions (IBR) and edge-based regions (EBR). IBRs are based on</p><p>　　extrema in intensity. Given a local intensity extremum, the brightness function along

32、rays emanating from the extremum is studied. This function itself exhibits an extremum at locations where the image intensity suddenly changes. Linking all points of the emanating rays corresponding to this extremum for

33、ms and IBR. EBRs are determined from corner points and edges nearby. Given a single corner point and walking along the edges in opposite directions with two more control points, a one-dimensiona</p><p>　　Ano

34、ther algorithm, termed Salient Region detector was proposed by Kadir et al. and is based on the probability density function (PDF) of intensity values computed over an elliptical region. For each pixel, the entropy extre

35、ma for an ellipse centered at this pixel is recorded over the ellipse parameter’s orientation, h, scale s and the ratio of major to minor axis k. From a sorted list of all region candidates the n most salient ones are ch

36、osen. For an extensive evaluation of a large number of af</p><p>　　Generally speaking, a descriptor is an abstract characterization of an image patch. Usually, the image patch is chosen to be the local envir

37、onment of an interest region. Based on various algorithms methods or transformations, the resulting character can be made rotation invariant or, at least partially, insensitive to affine transformations.</p><p

38、>　　Most approaches are based on gradient calculations or image brightness values. As a second part of the SIFT approach, Lowe proposed the use of descriptors based on stacked gradient histograms. The single histograms

39、 are calculated in a subdivided patch describe the gradient orientation in order to cover spatial information. Finally, they are concatenated to form a 128-dimensional descriptor. Recently Ke and Sukthankar, proposed the

40、 so called PCASIFT descriptor based on eigenspace analysis. They c</p><p><b>　　附錄B</b></p><p>　　基于DSP的通過(guò)局部特征實(shí)時(shí)物體識(shí)別嵌入式系統(tǒng)</p><p><b>　　摘要</b></p><p&g

41、t;　　在過(guò)去幾年中，對(duì)象識(shí)別已經(jīng)成為最熱門的任務(wù)，計(jì)算機(jī)視覺(jué)尤其是，這是推動(dòng)發(fā)展新的強(qiáng)大的算法，局部特征的物體識(shí)別。所謂'智能相機(jī)有足夠的權(quán)力分散的圖像處理變得越來(lái)越流行的各種任務(wù)，特別是在外地的監(jiān)視。它是一個(gè)非常重要的工具，強(qiáng)大的識(shí)別可疑車輛，人員或物體是否符合公眾安全。這只是局部識(shí)別功能的嵌入式平臺(tái)的基本功能。在我們的工作中，我們調(diào)查的任務(wù)是，目標(biāo)識(shí)別基于狀態(tài)最先進(jìn)的算法，在一個(gè)基于DSP的嵌入式系統(tǒng)。我們執(zhí)行一些功能強(qiáng)大

42、的算法識(shí)別物體，即有興趣點(diǎn)探測(cè)連同區(qū)域描述，并建立一個(gè)中型對(duì)象數(shù)據(jù)庫(kù)為基礎(chǔ)的詞匯樹，這是適合我們的專用硬件設(shè)置。我們仔細(xì)研究了該算法參數(shù)性能的嵌入式平臺(tái)。我們所研究的，國(guó)家最先進(jìn)的目標(biāo)識(shí)別算法，可以成功地部署在當(dāng)今智能相機(jī)，即使計(jì)算和內(nèi)存資源有嚴(yán)格的限制。</p><p>　　關(guān)鍵詞 數(shù)字信號(hào)處理；物體識(shí)別；本地功能；詞匯樹；</p><p><b>　　介紹</b>

43、</p><p>　　識(shí)別物體是一個(gè)最流行的任務(wù)領(lǐng)域中的計(jì)算機(jī)問(wèn)題。在過(guò)去十年中，大量科學(xué)工作者做出努力，建立強(qiáng)有力的目標(biāo)識(shí)別系統(tǒng)的外觀特征與局部特征的程度。對(duì)于這樣一個(gè)框架，以適用于現(xiàn)實(shí)世界中的幾個(gè)屬性是非常重要的：對(duì)旋轉(zhuǎn)不敏感，光照或觀點(diǎn)的變化，以及實(shí)時(shí)的行為和大規(guī)模行動(dòng)。目前的系統(tǒng)已經(jīng)有很多這些屬性，雖然不是所有的問(wèn)題已經(jīng)解決，但如今他們變得越來(lái)越有吸引力的行業(yè)列入產(chǎn)品的客戶市場(chǎng)。</p>&

44、lt;p>　　反過(guò)來(lái)，最近嵌入式視覺(jué)平臺(tái)，如智能相機(jī)已經(jīng)成功地出現(xiàn)了，不過(guò)，只有提供數(shù)量有限的計(jì)算和內(nèi)存資源。然而，嵌入式視覺(jué)系統(tǒng)已經(jīng)在我們的日常生活中。幾乎每個(gè)人的手機(jī)配備了攝像頭，因此可以被視為一個(gè)小型的嵌入式視覺(jué)系統(tǒng)。顯然，這會(huì)引起新的應(yīng)用程序，如導(dǎo)航工具，視障人士，或協(xié)作公眾監(jiān)督使用數(shù)以百萬(wàn)計(jì)的人造眼睛。此外，低價(jià)格的數(shù)字傳感器和需要增加公共場(chǎng)所的安全造成了巨大數(shù)量的增長(zhǎng)攝像機(jī)監(jiān)視用途。他們必須體積小，并處理大量的現(xiàn)有數(shù)據(jù)

45、的網(wǎng)站。此外，他們必須執(zhí)行專門的業(yè)務(wù)自動(dòng)和人機(jī)交互。不僅在該領(lǐng)域的監(jiān)視，而且在家庭領(lǐng)域的機(jī)器人，娛樂(lè)，軍事和工業(yè)機(jī)器人技術(shù)，嵌入式計(jì)算機(jī)視覺(jué)平臺(tái)越來(lái)越受歡迎，這應(yīng)歸功于其對(duì)環(huán)境的魯棒性逆境。特別是基于DSP的嵌入式平臺(tái)很受歡迎，因?yàn)樗鼈児δ軓?qiáng)大且廉價(jià)的處理器，仍然是小規(guī)模和效率方面的能耗。隨著數(shù)字信號(hào)處理器提供了最大的靈活性，軟件運(yùn)行，相對(duì)于其他嵌入式單位像臺(tái)塑作為，專用集成電路或芯片，其目前的成功并不奇怪。</p>&l

46、t;p>　　對(duì)于已經(jīng)提到的原因，認(rèn)識(shí)的任務(wù)是一個(gè)非常重要的研究領(lǐng)域。然而，在這方面的一些屬性的嵌入式平臺(tái)的嚴(yán)格限制的可行性目前的狀態(tài)，最先進(jìn)的方法。例如，可用的記憶體數(shù)量的設(shè)備上嚴(yán)格限制數(shù)量的對(duì)象在數(shù)據(jù)庫(kù)中。因此，建立一個(gè)嵌入式對(duì)象識(shí)別系統(tǒng)，一個(gè)目標(biāo)是使大量的數(shù)據(jù)來(lái)表示一個(gè)單獨(dú)的對(duì)象盡可能小，以便最大限度地發(fā)揮一些識(shí)別物體。另一個(gè)重要方面是實(shí)時(shí)這些系統(tǒng)的能力。算法必須足夠快將業(yè)務(wù)在現(xiàn)實(shí)世界中。他們必須健全和用戶友好的，否則，產(chǎn)品配

47、備了這種功能，只不過(guò)是沒(méi)有吸引力的潛在客戶。例如，在一個(gè)互動(dòng)參觀博物館，識(shí)別物體在移動(dòng)設(shè)備上，必須足夠快，以便連續(xù)性指導(dǎo)。正式地講，我們認(rèn)為這是一個(gè)應(yīng)用程序需要軟實(shí)時(shí)系統(tǒng)的行為。顯然，這只是一個(gè)例子，以及確切的含義實(shí)時(shí)取決于具體應(yīng)用。我們?nèi)匀徽J(rèn)為物體識(shí)別系統(tǒng)是實(shí)時(shí)的能力，如果能夠提供至少一個(gè)結(jié)果每秒。這已經(jīng)足以讓許多服務(wù)等應(yīng)用的例子，介紹了上述互動(dòng)博物館。但是，很顯然，這個(gè)定義不符合其他應(yīng)用程序，并改善吞吐量需要識(shí)別物體的幀速率，例如，

48、結(jié)合目標(biāo)跟蹤?？傊⒁粋€(gè)全功能的識(shí)別系統(tǒng)在嵌入式平臺(tái)原來(lái)是一個(gè)具有挑戰(zhàn)性的問(wèn)題給所有不同的方面和環(huán)境限制考慮。</p><p>　　在這項(xiàng)工作中，我們描述的一種方法來(lái)部署一個(gè)中型物體識(shí)別系統(tǒng)原型基于DSP的嵌入式平臺(tái)。以我們所知，我們是第一個(gè)廣泛的調(diào)查有關(guān)的問(wèn)題識(shí)別物體在嵌入式系統(tǒng);現(xiàn)在這是唯一的工作，學(xué)習(xí)的影響，各種參數(shù)對(duì)識(shí)別性能和運(yùn)行時(shí)行為。我們挑選了一套高層次的算法來(lái)描述物體的一套外觀特征。作為一個(gè)典型

49、的局部特征識(shí)別系統(tǒng)，我們使用不同的高斯（狗）關(guān)鍵點(diǎn)和主成分分析尺度不變特征變換（ PCASIFT ）描述建立緊湊對(duì)象的意見(jiàn)。安排這方面的信息的一個(gè)聰明treelike數(shù)據(jù)結(jié)構(gòu)基于K - means聚類，一個(gè)所謂的詞匯樹，實(shí)時(shí)行為實(shí)現(xiàn)。通過(guò)運(yùn)用專門的壓縮機(jī)制，大小的數(shù)據(jù)結(jié)構(gòu)可以進(jìn)行交易的對(duì)識(shí)別性能，從而準(zhǔn)確調(diào)諧的屬性識(shí)別系統(tǒng)對(duì)某一特定的硬件平臺(tái)就可以執(zhí)行。因?yàn)樗@示了廣泛的評(píng)價(jià)的同時(shí)考慮，特殊性能的算法和專用的特殊的硬件優(yōu)勢(shì)，大量增加識(shí)別

50、性能和吞吐量是可以實(shí)現(xiàn)的。</p><p>　　其余本文結(jié)構(gòu)如下，我們概述的事態(tài)發(fā)展在這兩個(gè)領(lǐng)域，我們正在把我們的工作。一方面，我們列出了一些參考的背景下目標(biāo)識(shí)別的計(jì)算機(jī)視覺(jué);另一方面，我們列舉了一些出版物，從該地區(qū)的嵌入式智能傳感器。詳細(xì)說(shuō)明方法參與建設(shè)我們的目標(biāo)識(shí)別算法是在第3部分。，我們大綱的框架，讓我們?cè)敿?xì)的培訓(xùn)和實(shí)施我們的系統(tǒng)。我們密切描述設(shè)計(jì)中的所有步驟的做法，讓方說(shuō)明的替代方法，我們?cè)u(píng)估我們的實(shí)驗(yàn)系

51、統(tǒng)的一個(gè)具有挑戰(zhàn)性的對(duì)象數(shù)據(jù)庫(kù)，并討論實(shí)時(shí)和現(xiàn)實(shí)世界的問(wèn)題。此外，我們調(diào)查的一些特殊功能的辦法，并闡明相依的若干參數(shù)對(duì)系統(tǒng)的整體性能。工作結(jié)束時(shí)的最后一些說(shuō)明和展望未來(lái)的工作。</p><p><b>　　相關(guān)工作</b></p><p>　　在下面我們將簡(jiǎn)要介紹該專題的局部特征的物體識(shí)別。由于大量的文獻(xiàn)資料，我們將集中于最有前途的方法利用當(dāng)?shù)靥攸c(diǎn)，是指那些有某種算法

52、與我們的工作。我們還將在短期簡(jiǎn)介識(shí)別物體在嵌入式系統(tǒng)，其中，由于稀疏的現(xiàn)有方法，包含全球和地方的方法，以及算法實(shí)現(xiàn)FPGA和DSP為基礎(chǔ)的平臺(tái)。</p><p>　　當(dāng)?shù)爻霈F(xiàn)的視覺(jué)識(shí)別物體成為受歡迎的發(fā)展后，強(qiáng)大的利益區(qū)域的探測(cè)器和描述。早期功能齊全的目標(biāo)識(shí)別系統(tǒng)處理所有個(gè)人算法步驟及有關(guān)問(wèn)題提出了施密德和莫爾和席勒和Crowley 。主要的想法基于局部特征識(shí)別物體保持對(duì)象申述收藏當(dāng)?shù)夭蓸诱f(shuō)明。換言之，部分地方出

53、現(xiàn)一個(gè)單一的對(duì)象編碼描述，和一套這些描述形式的最后對(duì)象的代表性。找到區(qū)分地區(qū)，所謂的利益區(qū)域探測(cè)器的使用，該區(qū)域或分找到的特殊視覺(jué)的獨(dú)特性。附近的這些地區(qū)是后來(lái)編碼使用一種特殊變換建立一個(gè)描述提供一些可取的本質(zhì)屬性。除了對(duì)光照變化不敏感，局部的觀點(diǎn)不變，申述套地方描述提供抗背景雜波和部分閉塞。不用說(shuō)，一個(gè)所謂的包描述符的代表性可以采用單一或數(shù)個(gè)不同的組合探測(cè)器和描述。</p><p>　　集體所有描述由多個(gè)對(duì)象（

54、即包描述符）是用來(lái)建立一個(gè)數(shù)據(jù)庫(kù)。鑒于這一新的數(shù)據(jù)庫(kù)和代表性的對(duì)象必須承認(rèn)，對(duì)應(yīng)計(jì)成投票計(jì)劃，以確定正確的匹配。確定這些書信是一項(xiàng)復(fù)雜的任務(wù)。描述高維特征向量和描述符匹配的查詢手段確定確切的近鄰在數(shù)據(jù)庫(kù)中。不幸的是，到目前為止，沒(méi)有任何演算法，都可以準(zhǔn)確近鄰點(diǎn)在高維空間，是比任何更有效的徹底搜查。由于大量的對(duì)象，大量的本地描述符，分別這種類型的信息管理是臃腫，效率低下。因此，許多不同的方法來(lái)近似的解決辦法的一條有效途徑已提議保持業(yè)績(jī)的總

55、體目標(biāo)識(shí)別系統(tǒng)的管理。</p><p>　　基本原則和地區(qū)的興趣點(diǎn)是尋找地點(diǎn)和地區(qū)的形象，展示一個(gè)預(yù)定義的財(cái)產(chǎn)使他們特別在其當(dāng)?shù)厣鐓^(qū)。這個(gè)屬性應(yīng)當(dāng)使該區(qū)域有別于其鄰居和探測(cè)反復(fù)。此外，這些功能的檢測(cè)應(yīng)以盡可能最好的，照明和觀點(diǎn)不變。</p><p>　　第一個(gè)重要的興趣點(diǎn)檢測(cè)器，即所謂的哈里斯角探測(cè)器，有人提議在1988年由Harris和斯蒂芬斯。它出色的重復(fù)性和展品后來(lái)用于識(shí)別物體的目的

56、施密德和莫爾。的擴(kuò)展，哈里斯探測(cè)器，包括大規(guī)模的信息后來(lái)被報(bào)道Mikolajczyk和施密德作為哈里斯一拉普拉斯探測(cè)器和使用的是Schaffalitzky和Zisserman formulti視圖匹配無(wú)序圖片集。另一種方法來(lái)檢測(cè)斑點(diǎn)狀結(jié)構(gòu)的圖像搜索點(diǎn)的行列式的Hessian矩陣假設(shè)當(dāng)?shù)貥O端嗯，這是所謂的黑森州探測(cè)器。進(jìn)一步發(fā)展包括仿射方差導(dǎo)致哈里斯仿射和黑森州仿射探測(cè)器提出Mikolajczyk ， Mikolajczyk和施密德。&l

57、t;/p><p>　　目前最流行的兩個(gè)部分的做法稱為尺度不變特征變換（上海對(duì)外貿(mào)易學(xué)院）是由勞文提出，其中的第一部分是一個(gè)興趣點(diǎn)檢測(cè)。狗探測(cè)器采取不同的高斯模糊圖像作為一個(gè)近似的規(guī)模正?；绽购褪褂卯?dāng)?shù)刈罡叩拇饛?fù)中尺度空間作為一項(xiàng)指標(biāo)的關(guān)鍵點(diǎn)。補(bǔ)充功能探測(cè)器，在最大限度地穩(wěn)定極值區(qū)域（ MSER ）探測(cè)器，是由麥塔斯等。?？傊?， MSER探測(cè)器搜尋區(qū)域是光明或黑暗的比其周圍的環(huán)境，即周圍的黑暗，反之亦然光明像素。

58、首先，像素的排序在升序或降序進(jìn)行排序的強(qiáng)度值，根據(jù)地區(qū)不同類型的檢測(cè)。像素陣列，比上一季度美聯(lián)儲(chǔ)進(jìn)入聯(lián)盟找到算法和樹形數(shù)據(jù)結(jié)構(gòu)形狀維持，而節(jié)點(diǎn)包含有關(guān)像素街道，以及有關(guān)強(qiáng)度值的關(guān)系。最后，節(jié)點(diǎn)滿足一套預(yù)先確定的標(biāo)準(zhǔn)要求的樹木遍歷算法。</p><p>　　兩個(gè)仿射協(xié)變區(qū)域探測(cè)器提出了Tuytelaars和Van Gool ，強(qiáng)度為基礎(chǔ)的區(qū)域（ IBR技術(shù)）和優(yōu)勢(shì)為基礎(chǔ)的區(qū)域（ EBR ） 。 IBRs是基于極值的

59、強(qiáng)度。鑒于當(dāng)?shù)氐膹?qiáng)度極值，亮度功能沿射線產(chǎn)生的極值研究。這個(gè)功能本身展品的極值在地方的形象強(qiáng)度突然變化。連接所有點(diǎn)所產(chǎn)生的射線相應(yīng)這個(gè)極值形式和IBR技術(shù)。 EBRs決心從角落點(diǎn)和邊緣附近。鑒于一個(gè)角落點(diǎn)和走邊有兩個(gè)相反的方向有更多的控制點(diǎn)，一維類平行介紹了利用角球本身和載體指著從角落的控制點(diǎn)。研究函數(shù)的紋理和使用額外的限制，一個(gè)平行四邊形是選定一個(gè)EBR 。</p><p>　　另一種算法，稱為凸區(qū)域探測(cè)器是由

60、卡迪爾等。并基于概率密度函數(shù)（ PDF ）的強(qiáng)度值計(jì)算的一個(gè)橢圓形的區(qū)域。每個(gè)像素，熵極值為橢圓中心在此記錄像素的橢圓參數(shù)的方向，氫，硫和規(guī)模的比例，主要以短軸光從排序名單的候選人的所有地區(qū)的N最突出的是選擇。對(duì)于一個(gè)廣泛的評(píng)估了大量仿射區(qū)域探測(cè)器提及的工作。</p><p>　　一般而言，一個(gè)描述符是一個(gè)抽象的表征圖像修補(bǔ)程序。通常情況下，圖像修補(bǔ)程序推選為當(dāng)?shù)丨h(huán)境的興趣區(qū)域。根據(jù)不同的算法，方法或轉(zhuǎn)變，由此產(chǎn)

61、生的性質(zhì)，可旋轉(zhuǎn)不變或至少部分敏感仿射變換。</p><p>　　大多數(shù)的做法是基于梯度計(jì)算或圖像的亮度值。作為第二部分的上海對(duì)外貿(mào)易學(xué)院的做法，羅威建議使用描述的基礎(chǔ)上疊加梯度直方圖。單直方圖計(jì)算在細(xì)分補(bǔ)丁描述梯度方向，以涵蓋空間信息。最后，它們級(jí)聯(lián)形成一個(gè)128的三維描述。最近柯和Sukthankar ，提出了所謂的PCASIFT描述特征空間分析的基礎(chǔ)上。他們計(jì)算的一個(gè)主要成分分析（ PCA ）的梯度特征空間

62、圖像，代表人數(shù)超過(guò)兩萬(wàn)形象補(bǔ)丁。該描述符的新形象瓦所產(chǎn)生的梯度投影的瓷磚上precalculated特征空間，只保留的D最重要的特征向量。因此，一個(gè)有效的壓縮描述維度實(shí)現(xiàn)， coevally保持業(yè)績(jī)的速度可比原來(lái)的上海對(duì)外貿(mào)易學(xué)院描述。密切相關(guān)的上海對(duì)外貿(mào)易學(xué)院的做法，梯度位置和方向直方圖（ GLOH ）描述是由Mikolajczyk和施密德。反對(duì)篩選梯度直方圖計(jì)算，細(xì)圓而不是粗糙的矩形網(wǎng)格，導(dǎo)致272二維直方圖。常設(shè)仲裁法院隨后用來(lái)降