EE678 Application Presentation Content Based Image Retrieval Using Wavelets

Transcription

1 EE678 Alication Presentation Content Based Image Retrieval Using Wavelets Grou Members: Megha Pandey iitb.ac.in 02d07006 Gaurav Boob 02d07008 Abstract: We focus here on an effective strategy that allows user to ose a visual query and retrieve a set of images from a database that satisfy his criteria of ictorial similarity without requiring any semantic exression of the query ictures. We discuss here Rotation Invariant wavelet acket transform aroach to this kind of Image retrieval system. Index Terms: Content Based Image Retrieval, Wavelet Packet Transform, Polar Transform I. INTRODUCTION: We are facing an exlosion of visual information as images are acquired, or generated in increasing number and at diminishing costs. As the size and the number of image assets grows, traditional methods of image archiving and retrieval become costly and inefficient in terms of the human assistance needed to erform this task. An Image Retrieval System is required to effectively and efficiently handle such huge amount of images. Such system will hel users to retrieve images based on their contents. There remain two basic roblems in this area: First one is the roblem of efficient and meaningful image segmentation where we break the image into low level arts like color, shae, texture and satial location. The second one is the vast ga that lies between these low level features and high level or semantic exressions contained in image like the image of car, house etc. Here we try to develo effective method to achieve a reasonable solution to these roblems. II. WAVELET PACKET TRANSFORM: Wavelet transform (WT) is a mathematical tool that can decomose a temoral signal into a summation of time-domain basis functions of various frequency resolutions. The continuous one-dimensional Wavelet transform is a decomosition of f(t) into a set of * basis functions Ψ ab, () t called wavelets. 1 t b Ψ a, b() t = Ψ ( ) a a where a is the scale arameter and b is the dilation arameter.

2 As continuous WT is redundant, the scale and translation arameters are also discretized to yield Discrete WT for which a dyadic lattice is often chosen (a = 2^m & b = n2^m). m? m 2 2 W[m,n]=2 f [ k] Ψ[2 k n ]? Effectively wavelet transform searates a signal into low frequency comonents (scaling coefficients) and high frequency comonents(wavelet coefficients).the WT alies the wavelet transform recursively to the low ass result. If it is to high ass results as well, we get wavelet acket transform. The wavelet acket transform can be viewed as a tree. The root of the tree is the original data set. The next level of the tree is the result of one ste of the wavelet transform. Subsequent levels in the tree are constructed by recursively alying the wavelet transform ste to the low and high ass filter results of the revious wavelet transform ste. III. CONTENT BASED IMAGE RETRIEVAL: In a tyical CBIR system, a user has an image and wants to find similar images from a large database. Tyically in a content based aroach, the images are retrieved directly based on their visual content such as color, texture, shae etc. A Content Based Image Retrieval system consists of three comonents: feature extraction, feature indexing and retrieval engine. In feature extraction comonent extracts visual feature information from the image database, the feature indexing comonent organizes the visual feature information to seed u the query rocessing and the retrieval engine rocesses the user query and rovides user interface. A large number of content based image retrieval systems have been built e.g. QBIC, VisualSeek and Photobook. In QBIC content based queries such as query by examle

3 image, query by sketch and drawing, query by selected color are suorted. In VisualSeek both content based and text based queried are suorted. In this reort we use feature based method. This is a two ste method. First, for each image in database, a feature vector is characterizing some image roerties are comuted using mathematical techniques and stored in feature database. Second, given a query image, its feature vector is obtained and then comared to other feature vectors already resent in the database and images more similar to query image are returned. Two main features that are used for image comarison are Color feature and Texture feature. The method being discussed here uses texture features for image retrieval. IV. ROTAION INVARIANT POLAR-WAVELET TEXTURE FEATURE FOR CBIR: These days most of the CBIR systems are based on image feature extraction and comarison. Most widely used image feature are color, shae, texture etc. Emhasis of Rotation Invariant Polar-Wavelet lies on texture feature of image. The feature extraction rocess involves a olar transform followed by an adative row shift invariant wavelet acket transform. To reduce feature dimensionally, only most dominant olar wavelet energy signature are selected. 1. TEXTURE FEATURE: Texture analysis is imortant for classification, segmentation or detection of image based on local satial atterns of intensity and color. Textures are relication and combination of various basic atterns. Most of the system which uses texture features for image comarison like QBIC, MR-SAR model, assumes image having same orientation. This assumtion is not valid in most of the ractical alications. The method resented here is indeendent of image orientation.

4 2. POLAR TRANSFORM: The olar transform converts an image into its equivalent image in olar form. Such a olar image is rotation invariant but is row shifted.the Polar Transform samles any image of NxN from 0 to 360 degrees, S times to roduce a Sx[N/2] image. Polar form (i,j) of the inut image f(x,y) can be comuted as follows: N 2πi N 2πi i (, j) = f({ } + { jcos( )},{ } { jsin( )} 2 S 2 S N for i = 0,..,S-1 and j = 0,.., 1 2 In the above figure, (a) is a samle texture, (d) its olar transform, (b) & (c) are images with anti-clockwise rotation of 60 and 120 degrees resectively, and (e) & (f) their resective olar transforms. Similarity of olar transforms or rotated versions can be easily seen. 3. ADAPTIVE ROW SHIFT INVARIANT WAVELET PACKET TRANSFORM: This transform emloys a air of quadrature mirror filters(qmf) to obtain orthonormal reresentation. Row shift invariance is achieved by building redundant set of wavelet acket coefficients for one additional circular shift. That is on each level we comute four eriodic images with no shift as follows: C 8k (i, j) = m n h(m) h(n) C k,(m+2i, n+2j) C 8k+1 (I, j) = m n h(m) g(n) C k,(m+2i, n+2j) C 8k+2 (I, j) = m n g(m) h(n) C k,(m+2i, n+2j) C 8k+3,(i, j) = m n g(m) g(n) C k,(m+2i, n+2j) where i = [N/2 ] 1 and j = [M/2 ] 1and C 0, (i, j) 0 = x( i, j) given by the intensity levels of image x at row i and column j.

5 These coefficients aear the same if C is circularly shifted by 0, 2, 4..., 2 n rows.similarly we comute another four eriodic images whose coefficients aear the same if C is shifted by 1, 3, 5,..., 2 n+1 rows. C 8k+4 (i, j) = m n h(m) h(n) C k,(m+2+1i, n+2j) C 8k+5 (I, j) = m n h(m) g(n) C k,(m+2+1i, n+2j) C 8k+6 (I, j) = m n g(m) h(n) C k,(m+2i+1, n+2j) C 8k+7,(i, j) = m n g(m) g(n) C k,(m+2i+1, n+2j) In this manner original image has been decomosed into eight quarter size images. In order to obtain the best basis reresentation we can adatively select some sub-bands instead of decomosing all of them. The criterion used is the information cost of the sub-bands. A sub-band is decomosed further only if its information cost is higher than the sum of all next level sub-bands. Hence the best basis can be determined by an efficient recursive selection rocess based on local minimization of the information cost function. By reeating this rocedure recursively to all levels, we can get wavelet acket coefficients for all circular row shifts in log N stes. 4. WAVELET ENERGY SIGNATURE: The generated wavelet transform coefficients are now shift-invariant; however, we need to reduce these large numbers of wavelet coefficients. We reduce feature dimensionality of the wavelet coefficient by comuting energy signature of each sub-bands generated by adative row shift wavelet acket transform. Thus we have equal number of energy signatures and sub-bands generated by adative row shift invariant wavelet acket transform. Out of these signatures, we choose only K most dominant energy signatures (with highest energy) as feature vector. 5. COMPARISON WITH QUERY IMAGE: Given a query image from the user, the roosed CBIR system extracts energy signature of the image using adative row-shift invariant wavelet acket transform. These energy signatures are then comared with the signatures of images resent in the database. The th similarity between query image and n image in database is defined by- q n M q n i i i = 0 d( f, f ) = f f q where f is feature vector of query image and database. f n is feature vector of image in V. RESULTS: The results obtained using this method are shown below in the order of decreasing level of similarity and increasing distance from query image.

6

7 VI. CONCLUDING REMARKS: A rotation invariant texture feature for image-retrieval was discussed. The method is quite efficient with only (n log n) comlexity. The method has been shown to achieve high retrieval accuracy; esecially in the cases where images of interest have different orientations. VII. REFRENCES: [1]. Rotation Invariant Texture Feature for Content Base Image Retrieval by Chi-Man Pun and Moon-Chuen Lee [2]. An Effective Content Based Visual Image Retrieval System by Shu-Ching Chen, Mei-ling Shyu and Borko Furht. [3]. An overview of CBIR techniques by Sagarmay Deb and Yanchun Zhang.