INTRODUCTIONProcedurePre-processingDatabase ConstructionImage RecognitionResult and discussionEE368 Digital Image Processing 2007 Project: Group 12 1Image Recognition Technique using Local Characteristics of Sub-sampled Images Group 12: Do, Hyungrok Abstract—An image recognition technique utilizing a database of image characteristics is introduced. This technique is different from eigenimage method which requires a large amount of information of training set images in terms of the size of each image and the number of images in the database. Especially, this technique is useful for recognizing images which have fixed shape and structure such as paintings and documents. In this study, images of 33 different classic paintings taken by a common camera phone are used to construct the database and a MATLAB code is written for image recognition. 66 different images of the 33 paintings are tested and approximately 80 % of them are recognized correctly. In the code, low pass filters for noise reduction, morphological operators such as dilation and majority filter for clear and smooth boundaries and Haralick corner detector to find characteristic points are used. To construct a database which consists of small size images, original images are trimmed to be a smaller image which contains only the region of interest. Furthermore, each image is sub-sampled to be a fixed size gray image which is 200 pixels by 200 pixels. Sub-sampling can reduce discrepancy of trimming position and angle/position of camera. Finally, using Haralick corner detector, 100 corner points which have large cornerness are selected per image. The points are positioned on a 200 by 200 binary image, which is a reference image in database. Some conventional image processing techniques are applied to an input image. The resulting image is also converted to a binary 200 pixels by 200 pixels image and compared with the 33 reference images in the database being shifted and warped. I. INTRODUCTION ECENTLY, image recognition techniques have been studied for many applications. Especially, those techniques are useful for computer based automated recognition systems and mobile camera phones. Among them, one of the most well-known techniques is the eigenimage method. Eigenimage method is capable of recognizing complex objects such as human face [1]. Furthermore, by collecting information from a large group of training set images possessing same characteristics, this novel algorithm can classify and recognize an enormous number of different groups. In the other word, if enough number of training set images can be obtained, then this intelligent algorithm can discriminate any sophisticated difference of images. However, if the number of groups is small, then the characteristics of images required to discriminate groups can be much simpler. Therefore, in this study, I tried to use the minimum amount of information for image recognition. In the next section, two different numbers of characteristic points per reference images in database are used and compared. Do, Hyungrok, Ph. D. Student in Mechanical Engineering Department of Stanford University, Bldg 520 room 520i Duena St. Stanford CA, 94305 (phone: 650-353-1243; e-mail: [email protected]) Generally speaking, perception of an object does not require whole detailed information of it, but only some characteristics which are discernable. However, as the number of discernable groups increases, the characteristics of each group should be more sophisticated to be recognized correctly. For example, at least 33 different characteristics (e.g. point) are required to recognize images from 33 different groups. Fortunately, an image which has complex patterns or structure has many kinds of characteristics. Among them, corner points or cornerness are selected as the representative characteristic points in this study. Sub-sampling can be used when the number of pixels in the reference images is much larger than the number of groups to be discriminated. Needless to say, information in the image would be lost by sub-sampling. However, if number of pixels in the sub-sampled image is still much larger than the number of groups, then the information in the sub-sampled image is usually enough to distinguish different groups. A crucial merit of sub-sampling, except the small data size, is reduction of the errors or discrepancy which can be frequently caused when the region of interest is trimmed and detected. For example, let’s assume that a region of interest is shifted by 10 pixels and the trimmed image is sub-sampled by a factor of 5. Then the discrepancy or error of position would be reduced to 2 pixels. Similarly, rotation or warping in a proper range can be neglected in the sub-sampled image. Furthermore, by dilating characteristic points (e.g. corner points) in the sub-sampled image, this kind of problems can be easily resolved in a larger range of unexpected shifting, rotating and warping. II. PROCEDURE A. Pre-processing The first step of the pre-processing is excluding gray wall region utilizing ratio of intensities of blue color over red color. If the ratio of a pixel is around one, one can assume that the pixel has gray tone color, although all of the intensities of red, green and blue colors should be the same to be gray. These pixels are set to be black. The other ratios (e.g. green over red, etc.) are also used to perform the same processing for comparison. REE368 Digital Image Processing 2007 Project: Group 12 2As shown in fig. 1., the wall region is completely black, although, unfortunately, some parts of paintings also become black. One can observe noisy regions near the bottom of the figure. However, the region of interest, the painting in the center, has relatively clear boundaries compared to the noisy region. To clarify and strengthen the boundary of the painting, the figure is dilated by 3 by 3 structuring element. Additionally, to reduce the noise out of the region of interest, a low pass filter is applied before thresholding. In the procedure, the resulting image (fig. 1.) should be a gray image. Therefore, red color is selected as a representative color. Average intensity of the three colors is also used for comparison. The second step is trimming the region of interest. Presumably, the picture in the center should be the dominant painting. Therefore, the algorithm probing the dominant painting is designed to start from the center point of the