The Automatic Detection and Recognition of Objective Parts in Real-World Building Image

Now, we are constructing a City-streets Building Database System. Particularly, we want to use this system to recognize real-world building images and provide the necessery information about the building objects. Because in building image there are many objects, and only the parts of building objects are necessary for recognition, we have developed a method by which all the parts of building objects are first detected using a statistical algorithm, and then the image features are extracted from the parts. Here, we suggest that the texture and color features are extracted respectively using "wavelet transform" on "wavelet transform" theory and color clustering method for recognition. In experiments, the realworld building images taken with a different angle, light, and scale on some extent from the image patterns of the database, are used to test the accuracy and efficacy of the method. This paper is organized as follows. The detection of the building parts using the statistical algorithm is discussed in section 2. The extraction of color and texture features are given in section 3. The experiment results are presented in section 4. Finally, we summarize this paper and presented the future research work. and color-clustering. At the same time, a method 2 The detection of the building parts called "vote" is used to decide the candidates for building object. Finally, we will ~ i v e some experiin real-world building image ment results to reveal the efficacy and accuracy of the proposed method in this paper. For detecting the parts of building objects, we first detect the points representing texture informal Introduction tion. At present, based on the development of multimedia technology, the multimedia systems for processing real-world information are being developed widely in the world. The systems are needed not only to be able to effectively manage the da ta related to real-world objects, but also to present the ways of automatically recognizing real-world objects wing compiit er vision technology. Considering these needs, we are constructing a City-street ~ u i l d i n g Database using the da ta realted to building objects and building images. Particularly, based on the vision indexes of the database constructed using building images, the database can be accessed using realworld building images. For realizing the objectives mentioned a.. above, this paper will propose a nlethod of automatically detecting and recognizing a11 the parts of building objects in building image. Here, the method is suggested to first detect all the parts of building objects using a statistical algorithm, and then the texture and color features are extracted respectively based 'Address: 7-22 1 Roppongi, Minato-ku, tokyo 106 Japan. E-mail: jhmasak. iis .u-tokyo.ac. jp '~ddress: 7-22-1 Roppongi, Minato-ku, tokyo 106 Japan. E-mail: sakauchiQsak. iis .u-tokyo. ac. jp 2.1 The detection of the points representing texture informa.tion According to the practical needs, here the Y color component representing the luminance channel of CIE-XYZ color space[l] and a 3 x 3 window are employed to detect the points representing texture information. If the Y component value of the point located a t the center of the window is the largest or the least within the window and the absolute difference value between the center point and other i t certain point in the window is larger or smaller than a specilized threshold, the center point is tletected as a point representing texture information(Fig.1). 2.2 Removing the noise points and recovering the points representing texture informat ion Although the points representing texture information can be detected correctly by the abovementioned method, there are still many noises in the points representing texture information. This will make great influence to detecting the parts of Figure 1: A building image and the points extracted using the method of detecting texture information Figure 3: Clustering peaks based on the distribution of the points locating in the particular windows 1 H1(i) = maxH1(i + j), j = -5, --., +5 P ( i ) = 0 otherwise Figure 2: A original image of the points exracted from building image and the image of the points smoothed using the templetes. building objects afterword. Therefore, it is necessary to remove these noise points and recover the points representing texture information. In this system, some 3 x 3 and 4 x 4 templates have been used to do these. A result example using this method is shown in Fig.2. 2.3 Deciding the ranges of building parts along horizontal direction Here, the ranges of building parts along horizontal direction are detected using a statistical algorithm. The algorithm are shown as follows: 1. Compute the histogram of the points representing texture informition distributing along the direction of horizontal axis. For finding easily the really variance positions, it is necessary to smooth the histogram. Here, a t every position of the histogram, the average values of five-neighborhood is calculated to be the new values of the histogram as defined in equation (1). 2. Find all the positions of peaks ( fiveneighborhood maximum ) using equation (2). 3. Cluster these peaks. Here, a t every position of the peaks, a window with the size of 50 x 50 pixel is put up in the middle of image for clustering(Fig.3). The window is further divided into four small 25 x 25 sub-windows, and the proportions of the number of points locating in the 50 x 50 window to the number of the points locating in every the sub-window, are calculated respectively using equation(3). These four proportions are used to calculate all the distances of similarity between a certain peak and the right one from left to right by equation (4) for peaks clustering. If a distance is less than a specialized threshold, the corresponding two peak positions are considered belonging to the same building object. C"," Ti(u, v) = [E,,, @(x, Y ) + 0.51 (3) Here, (u,v) E window locating in peak i. (x, y) E the nth sub-window of the window locating a t peak i. Ti(u, v) equals one if the point locating a t (u,v) is the one representing the texture information, otherwise zero. P,!, : The proportion of the number of texture information points held by the window locating a t the position of peak i to the one held by its nth sub-window. Here n=1,2,3,4. 4. If a certain peak and its right first-neighboring peak do not belong to the same building object, the distance between it and its right secondneighboring peak is calculated as step 3. If the distance is less than a specialized threshold, and the distance between it and its first-neighboring one is less than a specialized threshold too, these three peak positions are regarded belonging to the same building object. 5. Decide the ranges of building parts. For doing this, here if the number of the peaks belonging to a certain cluster is one, the specialized neighboring range centered round the peak position is considered to be the range. Otherwise the range between the first peak and the last peak belonging to the cluster is considered as