W.Y.Ma, B.S.Manjunath. Texture features and learning similarity. IEEE CVPR, 1996.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....previous work, annotation was often regarded as a Boolean vector. In [8] and [10] normalized Hamming distance was used to combine the influence of the annotation and acted as a new feature for the retrieval. When the database is partially annotated and the annotations are used for learning, as in [11], neural networks are often used to train the similarity measurement. In our system, each model has a list of attribute probabilities, including the query model the user provides. If the query model is chosen from the database, we already have this probability list. This is the normal case as ....

W. Y. Ma and B. S. Manjunath, "Texture features and learning similarity, " Proc. IEEE CSCVPR, pp. 425--430, 1996.

....moments, i.e. T S terms, and the coupling term for the same moments on two consecutive scales, i.e. T (S 1) terms. Selecting these invariants, the total size of the rotation invariant descriptor is 4 T S T , which is smaller (if K 4) than the non rotational Gabor texture descriptor of [10], which has K T S values. 4.6 Similarity Assessment The texture descriptor is used in a content based image retrieval system similar to [5, 14] using the query by example paradigm. The system explores similarity assessment of image descriptors, but does not classify images by learning or ....

....were extracted. Therefore, the resulting texture database DB II consisted of 5,000 images. In this database each of the used query images had 39 corresponding images. Four di erent texture descriptors were implemented and tested. First, the gabor based non rotation invariant descriptor (GM) [10] with parameters K = 6; S = 4; T = 2 which yields a descriptor consisting of 48 values. The second descriptor was the rotation invariant descriptor (GI) with 30 values. Thirdly, a rotation invariant descriptor without the coupling terms M 3 (GIN) was exploit that consists of 16 values. Fourthly, a ....

W.Y. Ma and B.S. Manjunath. Texture features and learning similarity. In IEEE Conf. on Computer Vision and Pattern Recognition, pages 425{ 430, June 1996.

....the variation of local appearance as the imaging conditions change. Instead we account for the change in appearance with viewing illumination directions globally by populating the feature space with feature vectors from various sampled BTF s. As in many approaches in texture literature [18] 1] 23][19], we cluster the feature space to determine the set of prototypes among the population. Specifically, we invoke k means algorithm, which is based on the first order statistics of data, and finds a predefined number of centers in the data space, while guaranteeing that the sum of squared distances ....

W. Y. Ma and B. S. Manjunath. Texture features and learning similarity. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 425--430, 1996.

.... measurements are not necessarily consistent with human s perception [12, 15] This problem leads to retrieval results that do not always meet the users expectations [19] To improve retrieval performance, relevance feedback technique is often used to tune computational similarity measurements [6, 13, 14, 20, 22, 26]. Typically, each new query resets the similarity measurement back to its initial state, which is not perceptually consistent. Subsequent feedbacks for the query are used to adjust the weighting factors of the similarity measurement to improve retrieval performance. The main diculty with this ....

....invariant and perceptual mapping method. 4.2 Comparison of Retrieval Performance An experiment was conducted to compare the retrieval performance of various computational texture models. Among the various texture features available, Gabor features are very widely used for image retrieval (e.g. [13, 10, 21]) and have been reported to best match the results of human vision study [21] Therefore, Gabor features were used in the experiment. The following computational models of texture similarity were compared: Euclidean distance: This is the most commonly used dissimilarity measurement. Fuzzy ....

[Article contains additional citation context not shown here]

W. Y. Ma and B. S. Manjunath. Texture features and learning similarity. In Proc. IEEE CVPR '96, pages 425-430, 1996.

....fi b j (a 1 Delta b 1 ; a 2 Delta b 2 ) From (5.8) one obtains the frequency bandwidth B: we denote ff = q ln(2) 2 ) B = log 2 F oe 1 ff F oe 1 Gamma ff (5. 27) The angular bandwidth Omega Gamma which describes the wavelets width along the angular axis, can be defined [40] as the angle between the two lines, passing through the origin and tangent to the curve of half maxima (Fig. 5.5.a) The following 72 CHAPTER 5. ROTATION INVARIANCE a) b) Figure 5.4: a: The Cauchy wavelet (B=1.0, Omega =30 ffi , 120 ffi ) in the frequency domain. b: the same in spatial ....

W. Y. Ma and B. S. Manjunath. Texture features and learning similarity. In Proc. IEEE Computer Vision and Pattern Recognition Conference, pages 425--430, June 1996.

....in large image databases such as the query by image content (QBIC) system [4] Texture segmentation involves the identification of uniform textured regions in an image. Many techniques have This research was supported in part by NSF CDA 9711582. been used for the analysis of textures such as [5, 6]. With the restriction to a set of known textures, retrieval and segmentation problems are essentially reduced to a supervised classification task, which is amenable for standard techniques from pattern recognition and statistics. Techniques used for image segmentation include simple statistical ....

W. Y. Ma and B. S. Manjunath, "Texture Features and Learning Similarity" Proc. IEEE International Conference on Computer Vision and Pattern Recognition, San Francisco, CA, pp. 425-430, June 1996.

.... existing computational similarity measures are not perceptually consistent [3, 6] This problem leads to retrieval results that do not always meet the users expectations [7] To improve retrieval performance, relevance feedback technique is often used to tune computational similarity measures [4, 5, 8]. Typically, each new query resets the similarity measure back to its initial state, which is not perceptually consistent. Subsequent feedback for the query are used to adjust the weighting factors of the similarity measure to improve retrieval performance. The main di#culty with this method is ....

W. Y. Ma and B. S. Manjunath. Texture features and learning similarity. In Proc. IEEE CVPR '96, pages 425--430, 1996.

....can appear in the images in varying scales and orientations. Computational texture features commonly used for image retrieval include the statistical features of Tamura et al. 13] used in IBM s QBIC system, DFT in [12] and the Wold model [6] MRSAR in Wold model, Gabor features in [14] NeTra [9], and invariant texture space [5] Among these models, 5, 12, 14] are scale and orientation invariant. However, our studies [8] show that texture similarity computed based on these methods are not necessarily consistent with human s perception. On the other hand, Santini and Jain s fuzzy features ....

....3 and 4 than in conditions 1 and 2 because there are more training samples in the former conditions than in the latter. 4. APPLICATION EXAMPLE To further assess the invariant network s performance, texture image retrieval tests were performed. In a similar test conducted by Ma and Manjunath [9], human subjects were asked to classify similar Brodatz textures into the same group. During retrieval, textures in the same group as the query are considered as relevant. In our tests, texture groupings were derived from k means clustering of the textures in PTS. Out of the 60 textures in PTS, ....

W. Y. Ma and B. S. Manjunath. Texture features and learning similarity. In Proc. CVPR, pages 1160--1169, 1996.

.... focus in CBIR has moved to interactive systems and human in the loop that involves a human as part of the retrieval process [1, 2, 3, 4] Examples include interactive region segmentation [5] interactive image database annotation [2, 6] usage of supervised learning before the retrieval [7, 8], and interactive integration of keywords and high level concepts to enhance image retrieval performance [9, 10] Following the current research achievements, visual data in a database can be considered to contain feature vectors representing the content of the data. A feature vector represents ....

W. Y. Ma and B. S. Manjunath, "Texture features and learning similarity", in Proc. IEEE Conf. Comput. Vision Pattern Recognition, pp. 425-430, 1996

....DFB metric for texture image retrieval. Each of the ### # ### images is divided into ## overlapping sub images of ### # ### pixels, centered on a # # # grid over the original image. The first ## sub images are used as the training set and the last ## for retrieval. This kind division is similar to [11]. Thus a database of #### texture images is formed for learning, and another #### texture images for testing the retrieval performance. Texture features are calculated by using the Gabor filter banks as in [4] with four scales and six orientations. Applying these Gabor filters to an image ....

....In previous Section, we evaluate the DFB metric for texture image retrieval. The goal is to retrieve the same class images on the top matches. However, there are many visually similar textures (but usually in different classes) can not be retrieved even on the top 100 matches. Ma and Manjunath [11] discussed this problem and used the Learning Vector Quantization to learn similarity. Our DFB metric can be extended naturally to solve this problem. The basic idea is 20 30 40 50 60 70 80 90 100 0.78 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 Number of top matches considered ....

W. Y. Ma and B. S. Manjunath, "Texture features and learning similarity," Proc. CVPR, pp. 425--430, 1996.

....dramatically decreasing the weights of the links between images and keywords. In this paper, we decide to make hidden annotation a preprocessing stage, referred to as the learning stage, before any user can use the system. Learning similarities before retrieval is not a new idea. Ma and Manjunath [13] used a hybrid neural network to learn the similarities between objects by clustering them in the low level feature space. Parts of the data were used for training, and the others were used for testing. The second observation we have is that most of existing systems using hidden annotation either ....

....work, annotation was often regarded as a Boolean vector. In [11] and [12] normalized Hamming distance was used to combine the influence of the annotation and acted as a new feature for the retrieval. When the database is partially annotated and the annotations are used for learning, as in [13], neural networks are often used to train the similarity measurement. In our system, each model has a list of probabilities of having the attributes, including the query model the user provides. If the query model is chosen from the database, we already have this probability list. If the query ....

W. Y. Ma and B. S. Manjunath, "Texture Features and Learning Similarity", Computer Vision and Pattern Recognition, 1996. Proceedings CVPR '96, 1996 IEEE Computer Society Conference on, pp. 425-430, 1996.

....in images of natural scenes, and (2) normalizing the intensity, contrast, scale, and orientation of the textures used in the psychological experiment. It is well known, for example, that human s sensitivity of perceiving spatial frequency (i.e. spatial patterns) is dependent on contrast [Schiffman, 1996] . Therefore, these confounding factors should be removed from the experiment. 3.2 Texture Image Preparation A rich set of texture images are collected to construct the perceptual texture space (PTS) Fifty texture images are selected from the Brodatz album. The remaining images in the album are ....

....1978] are used in IBM s QBIC system. The Wold model [Liu and Picard, 1996] based on Fourier transform (spectral model) and Multiresolution Simultaneous Autoregressive model (MRSAR, statistical model) is used in MIT s PhotoBook. Gabor features (spectral model) are used in Ma and Manjunath s NeTra [Ma and Manjunath, 1996] and Leow and Lai s invariant texture space [Leow and Lai, 2000] The above features are easy to compute, but the texture similarity computed based on these features are not necessarily consistent to human s perception. On the other hand, Santini and Jain developed the fuzzy features contrast ....

W. Y. Ma and B. S. Manjunath. Texture features and learning similarity. In Proc. of IEEE Conf. CVPR, pages 1160--1169, 1996.

.... (CBIR) The overall motive of this paper is to outline a framework that is capable of performing separate higher level and lower level queries, and in addition, a methodology for integrating higher level and lower level approaches in CBIR, which traditionally have largely been treated separately [1, 2, 4], for the retrieval of images containing large manmade objects (e.g. architectural objects or buildings, etc. The goal This research was supported in part by the Army Research Oce under contracts DAAG55 98 1 0230 and DAAD19 99 10012 (John Hopkins University subcontract agreement 890548168) ....

....the total number of scales, and x and y are the rotated coordinates: x = x cos y sin ; y = x sin y cos (5) where = n N is the orientation. The scale factor a m ensures that the lter energy is independent of m. The values of a, u and v are calculated as described in [2, 4]. Type Total (T ) Retrieved (R) Correct (C) Recall Precision (C=T ) C=R) Global higher level analysis only 45 43 36 80.00 83.72 Global lower level analysis only (without ROI) 45 27 15 33.33 55.56 Global lower level analysis only (with ROI) 45 46 23 51.11 50.00 Integrated (Re nement) ....

W. Y. Ma and B. S. Manjunath, \Texture features and learning similarity," in IEEE International Conference on Computer Vision and Pattern Recognition, pp. 425-430, 1996.

No context found.

W.Y.Ma, B.S.Manjunath. Texture features and learning similarity. IEEE CVPR, 1996.

No context found.

W.Y. Ma and B.S. Manjunath, "Texture features and learning similarity," Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Francisco, pp. 425-430, June, 1996.

No context found.

W.Y.Ma and B.S.Manjunath,Texture features and learning similarity, Proc. IEEE International Conference on Computer Vision and Pattern Recognition, San Francisco, CA, pp. 425-430, June 1996

No context found.

W. Ma, B. Manjunath, Texture features and learning similarity, in: Proceedings of the 1996.

No context found.

W. Ma and B. Manjunath, "Texture features and learning similarity", IEEE CVPR, 1996.

No context found.

W. Y. Ma and B. S. Manjunath, "Texture Features and Learning Similarity", Computer Vision and Pattern Recognition, 1996.

No context found.

W. Y. Ma and B. S. Manjunath, "Texture Features and Learning Similarity", Computer Vision and Pattern Recognition, 1996.

No context found.

W Y Ma and B S Manjunath. Texture features and learning similarity. In Proc. IEEE Conf. Computer Vision and Pattern Recognition, pages 425--430, 1996.

No context found.

W. Y. Ma and B. S. Manjunath, "Texture Features and Learning Similarity", IEEE Proceedings CVPR '96, pp. 425-430, 1996.

No context found.

W. Y. Ma and B. S. Manjunath, "Texture features and learn- ing similarity," in Proc. IEEE Cony. on Computer Vision and Pattern Recognition, 1996, pp. 425-430.

No context found.

W. Y. Ma and B. S. Manjunath, "Texture features and learning similarity", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 425-430, San Francisco, CA, 1996.

No context found.

W. Y. Ma and B. S. Manjunath, "Texture features and learning similarity", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 425-430, San Francisco, CA, 1996.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC