A Study On The Algorithms For Content-based Image Feature Extraction

With the fast development of the multimedia technology and theachievements of the computer networks, the capacity of the digital image isincreasing incredibly. Much information is included in the images. But theseimages located nearly all over the world, and information it contains are notindexed. Thus, it requires a prompt and accurate method to find out theproper images. That is called the Image Retrieval technology, when we talkabout image retrieval in the computer science field, which meansContent-based Image Retrieval (CBIR), in contra to those of Text-basedImage Retrieval (TBIR).CBIR means extracting features from an image according to its content.The main idea is to extract the information of color, texture, shape, relativeposition of objects, etc. from an image, and build some propermulti-dimensional feature vectors as index to present the specific image. Ifothers inquiring an image, the CBIR system calculates the similarity of thefeature vectors, and returns the most similar images. The core matter ofimage retrieval is how to present the key feature of an image efficiently andcorrectly.As one of the most important vision feature of image, color is often used todescribe image contents and used widely. This thesis discussed the keyquestions on the usages of the characteristics of color, i.e., expressing colorand obtaining the characteristics of color.With the feature discussion of practical color-spaces such as RGB, YUV,HSV, CIE-XYZ and CIE-L*a*b*, the author paid much interest in evencolor-spaces, and gave the transform formulas between these color-spaces.The common ways to express the feature of color-space is colorhistogram, color moment, color set. And the similarity between colorfeatures can be measured by the methods that follow: color histogramintersection method, L1 distance method, L2 distance method, quadraticform distance method, Mahalanobis distance method.In order to evaluate the methods above, the author introduced theMPEG-7 recommendation's Average Recall (AR) and Average NormalizedModified Retrieval Rank (ANMRR), and gave the formulas within.Converting a image in RGB color-space into YUV color-space is easilyachieved. But the statistic result suggests that YUV color-space is not apractical one for Image Retrieval purpose with AR=83% and ANMRR=0.26with histogram intersection method, AR=79% and ANMRR=0.29 with L2distance method.Unlike RGB or YUV color-space, HSV color-space is even to humanvisual characteristics, that means HSV color-space is suitable for imageretrieval when focused on color feature. Taking advance of human's feelingto color, it quantifies the color set with unequal interval, and getcharacteristic vector, regarded as the color feature of image. The statisticresult shows that with 256 bin histogram, intersection method resultsAR=100%, ANMRR=0.01, L2 distance method results in AR=86%,ANMRR=0.13;and with 22 bin histogram, histogram intersection results inAR= 98% and ANMRR=0.06, L2 distance method results in AR=94% andANMRR=0.12. The result above means that low vectors dimensions resultslow average recall, and some images will not be covered.XYZ color-space is not even, but is a cube of 1*1*1, and the variationpattern is regular, so the author divided the cube into 216 or 125 or 64 or 27small cubes as partials to evaluate the image harshly. The result suggeststhat even on the best condition, image retrieval based on XYZ color-space isworse that HSV in 22 bins.L*a*b* is a even color-space with irregular shape, so it can not be directlyperceived through the senses. The author only provided a simple colordifference evaluation method, and resulted in AR=81%, ANMRR=0.251,nearly the worst retrieval method.It is natural that image under observation is divided into foreground objectand background by human vision. The image retrieval performanceevaluated mainly with the object-based measurement. Considering theimage characteristics, the author divided the image into 9 equal rectangles,and extracts the middle block with weight. By comparing with methodsintroduced above, the color image retrieval accords with the better humanvision perception. The improvement of retrieval efficiency is proved byexperiments. But the weight of the middle block should be given manually,and no automatic solution is found.As a result, when retrieving image according to color-spaces, HSVcolor-space is better than XYZ color-space, and XYZ is better than YUV andL*a*b*. While YUV and L*a*b* is nearly unpractical for technology purpose.On the contrary, image retrieval based on HSV color-space presents a goodadvantage.Image retrieval based on color-space keeps the image with only colorinformation, and lost other features such as texture, shape, position and soon. So color-space based image retrieval usually used to find images likesceneries, perspective view, and not suitable for small shapes.The author implemented a color-space image retrieving system as thebed for extracting feature algorithms. The system is developed on IBM PC,Windows XP Professional and programmed with Borland Delphi.This thesis holds certain referential value and practical significance inpromoting the development of retrieval technique of image database.