J. Weng, N. Ahuja, and T. Huang, "Learning recognition and segmentation using the cresceptron," International Journal of Computer Vision 25(2), pp. 109-- 43, 1997.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....and lots of different techniques have been reported. Those techniques include geometric feature matching [19, 20, 22] template matching [22] deformable template matching [26] Karhunen Lo eve Transformation (KLT) 14] singular value decomposition (SVD) isodensity lines [5] neural network [28, 29], etc. Although high recognition rate has been obtained in some of these works, most of the authors assume that faces have already been segmented or deal only with face images in an uniform background. Moreover, few authors [14] have considered the issue of recognizing faces from video ....

....the principal component analysis (PCA) of the cross product matrix and reconstructing the faces as a weighted sum of eigenvectors. Therefore, the same result are obtained by two approaches: KLT [14] of the statistical approach and linear autoassociator of neural network approach [28] In [29], Huang et al. proposed an neural network structure called Creceptron. The Creceptron is based on the idea of Fukushima s Neocognitron [31, 32] Specifically, the Creceptron uses a hierarchical networks structure that can grow automatically, adaptively and incrementally through learning. With ....

[Article contains additional citation context not shown here]

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the creceptron," in Proc. International Conference on Computer Vision, pp. 121--128, May 1993.

.... to perform feature selection (see [1] for a brief survey) The most prominent of these are: hand crafted systems [2] these are not suitable for general example based queries) hierarchical systems [6] 4] where a priori processing is needed to structure the database; self organizing approaches [8], where the system continuously adapts itself to examples; Gaussian Mixture Model (GMM) based systems [5] where the class distributions are approximated using GMMs and then the discriminative power of the features is determined using KullbackLeibler Divergence. In the CBIR system in [3] where ....

....constrained to have linear time complexity w.r.t. number of features, therefore the feature selector should assume only linear discriminance based classifiers. 6. The module should be able to handle example sets of sizes as small as 5. As can be seen, 2] 6] 4] do not meet requirement 4, [8] is not designed for the large number of features involved, and [5] cannot handle the small example set sizes which are available. Our present work addresses the feature selection problem in the context of the above mentioned requirements. One important fact to note at this point is the difference ....

J. Wang, N. Ahuja, and T. S. Huang. Learning recognition and segmentation using the cresceptron. In Proc. Int. Conf. on Computer Vision, Berlin, Germany, pages 121--128, May 1993.

.... proposed over the years (Pitts McCullough, 1947; Fukushima, Miyake, Ito, 1983; Sandon Urh, 1988; Zemel, Mozer, Hinton, 1990; Le Cun et al. 1990; Mozer, 1991; Swain Ballard, 1991; Hummel Biederman, 1992; Lades et al. 1993; Califano Mohan, 1994; Schiele Crowley, 1996; Mel, 1997; Weng, Ahuja, Huang, 1997; Lang Seitz, 1997; Wallis Rolls, 1997; Edelman Duvdevani Bar, 1997) Many of these neurally inspired systems involve constructing banks of feature detectors, often called receptive fields (RFs) each of which is sensitive to some localized spatial configuration of image cues but invariant ....

.... Duvdevani Bar, 1997) Minimizing Binding Errors 249 intermediate network layers (Fukushima et al. 1983) or involving purely unsupervised learning principles that home in on features that occur frequently in target objects (Fukushima et al. 1983; Zemel et al. 1990; Wallis Rolls, 1997; Weng et al. 1997). While architectures of this general type have performed well in a variety of difficult, though limited, recognition problems, it has yet to be proved that a few stages of simple feedforward filtering operations can explain the remarkable recognition and classification capacities of the human ....

Weng, J., Ahuja, N., & Huang, T. S. (1997). Learning recognition and segmentation using the cresceptron. Int. J. Comp. Vis., 25(2), 109--143.

....effect of using a tree as opposed to a flat eigenspace. 1 Introduction An alternative to hand crafting features is the selforganizing approach, in which the machine will automatically determine what features to use and how to organize the knowledge structure such as the work of the Cresceptron [12] and eigenfaces [11] for view based recognition. Allowing the system to organize itself, however, raises some important efficiency issues. The first one is the feature selection issue. If the distribution is known, adding more features always produces better results (or at least not worse results) ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the Cresceptron, " in Proc. International Conference on Computer Vision, pp. 121--128, May 1993. Berlin, Germany.

....are updated by a supervised learning method. After training, the neurons in the higher layer have the tendency to respond selectively to some complicated features despite the variance in the same feature samples. Weng et al. in 1996 proposed a dynamic neural network model called Cresceptron [8], which could automatically grow a hierarchical of maps directly from image input by a learning with teacher process. The network grows by creating new neurons, connection and architecture which memorize new image structures and context as they are detected. Although these studies address global ....

J. Weng, N. Ahuja, and T. Huang, "Learning recognition and segmentation using the cresceptron," International Journal of Computer Vision 25(2), pp. 109-- 43, 1997.

....distance will be useless in differentiating a car from a fire hydrant. An alternative to hand crafting features is the self organizing approach, in which the machine will automatically derive what features to use and how to organize the knowledge structure such as the work of the Cresceptron [6] and eigenfaces [7] for view based recognition. In this framework, the recognition phase of the system is preceded by a learning phase. The learning phase focuses on the methods by which the system can automatically organize itself for the task of object recognition, giving it a wide range of ....

....position, and orientation of the objects in the images is allowed. The automatic selection of well framed images is an unsolved problem in general. Techniques have been proposed to produce these types of images, using, for example, pixel to pixel search [7] hierarchical coarse to fine search [6], or genetic algorithm search [30] This reliance on wellframed images is a limitation of the work; however, there are application domains where this limitation is not overly intrusive. In image databases, for example, the human operator will pre process the image data for objects to store in the ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the Cresceptron," in Proc. International Conference on Computer Vision, pp. 121--128, May 1993. Berlin, Germany.

....will probably be useless for grading lumber. An alternative to hand crafting features is the selforganizing approach, in which the machine will automatically determine what features to use and how to organize the knowledge structure. This approach can deal directly with complex, real world images [16] because the system is able to learn the environment in which it is to function, thereby making it much more flexible than its static counterparts. Self organizing image retrieval systems are open ended, allowing them to learn and improve continuously. As such, these systems are not restricted by ....

....attention, image scanning, and foviation technique. The visual attention mechanism provides a well framed portion of the image (i.e. a subimage) that contains a single object of interest in a standard position and scale. The image scanning technique could be manual selection, simple scanning [16] or sophisticated methods [13] The foviation procedure maps any image to a standard sized fovea. The details of this visual attention mechanism is beyond the scope of this paper. 2 Self Organizing, Hierarchical Optimal Subspace Learning and Inference Framework (SHOSLIF) The SHOSLIF [14] uses ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the cresceptron," in Proc. International Conference on Computer Vision, pp. 121--128, May 1993.Berlin, Germany.

....training with momentum. More sophisticated and more powerful training methods (such as the statistical methods) have been used when the input dimention is high and the number of classes is large. A different example of automatic feature derivation for face recognition is the Cresceptron [Weng et al. 1993] tested for general object segmentation and recognition, including faces. The method uses multilevel retinotopic layers of neurons to automatically determine the configuration of its network in the training phase. Unlike most neural networks, the Cresceptron does not use back propagation for ....

Weng, J., Ahuja, N., and Huang, T. S. (1993). Learning recognition and segmentation using the Cresceptron. In Proc. International Conference on Computer Vision, pages 121--128. Berlin, Germany.

....phase of the system is preceded by a learning phase. The learning phase focuses on the methods by which the system can automatically organize itself for the task of object recognition, giving it a wide range of generality. This approach can deal directly with complex, real world images [33] because the system is general and adaptive, thereby making it much more flexible than its static counterparts. Self organizing image retrieval systems are open ended, allowing them to learn and improve continuously. As such, these systems are not restricted by their human designer s limitations ....

....of the image that contains a single object of interest in a standard position and scale. During the training phase, the attention mechanism finds areas of interest using a scanning technique. The scanning is accomplished using a manual selection, a simple scanning approach such as discussed in [33], or a more sophisticated technique such as described in [29] The foviation procedure maps this area of interest to a standard sized fovea. Because the pixels nearer the center of this fovea are more important than those nearer the edges, the pixels in the fovea are weighted appropriately. The ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the Cresceptron," in Proc. International Conference on Computer Vision, pp. 121--128, May 1993.Berlin, Germany.

....an alternative to handcrafting the features. In this approach, a learning phase focuses on the methods by which the system can automatically organize itself for the task of image retrieval, giving it a wide range of generality [13] This approach can deal directly with complex, real world images [17] because the system is able to learn the image environment in which it is to function, thereby making it much more flexible than its static counterparts. Selforganizing retrieval systems are open ended, allowing them to learn and improve continuously. As such, these systems are not restricted by ....

....the visual saccades learned in the training phase of the system. Each time the Tree is invoked, it uses a fovea image extracted from the original image. These fovea images are assumed to be well framed images of objects that are extracted using a scanning technique such as the one described in [17] or [14] so that the size and position of an object in the fovea image are normalized. When a node in the Space Tessellation tree is activated, the input subimage is treated as a mn dimensional vector. The Karhunen Lo eve projection [8] 10] is used on this image vector to select the Most ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the creceptron, " in Proc. International Conference on Computer Vision, pp. 121--128, May 1993.Berlin, Germany.

....distance will be useless in differentiating a car from a fire hydrant. An alternative to hand crafting features is the self organizing approach, in which the machine will automatically determine what features to use and how to organize the knowledge structure such as the work of the Cresceptron [44] and eigenfaces [43] for view based recognition. In this framework, the recognition phase of the system is preceded by a learning phase. The learning phase focuses on the methods by which the system can automatically organize itself for the task of object recognition, giving it a wide range of ....

....position, and orientation of the objects in the images is allowed. The automatic selection of well framed images is an unsolved problem in general. Techniques have been proposed to produce these types of images, using, for example, pixel to pixel search [43] hierarchical coarse to fine search [44], or genetic algorithm search [41] This reliance on wellframed images is a limitation of the work; however, there are application domains where this limitation is not overly intrusive. In image databases, for example, the human operator will pre process the image data for objects to store in the ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the Cresceptron," in Proc. International Conference on Computer Vision, pp. 121--128, May 1993.Berlin, Germany.

....of the system is at the signal level instead of at the knowledge (e.g. shape) level. In this type of framework, a training phase finds salient features to use in the subsequent recognition phase of the system. These types of approaches can deal directly with complex, real world images [14] 20] [21] because the system is general and adaptive. The efficient selection of good features, however, is an important issue to consider [2] A well known problem in pattern recognition is called the curse of dimensionality more features do not necessarily imply a better classification success rate. ....

....position, and orientation of the objects in the images is allowed. The automatic selection of well framed images is an unsolved problem in general. Techniques have been proposed to produce these types of images, using, for example, pixel to pixel search [20] hierarchical coarse to fine search [21], or genetic algorithm search [18] This reliance on wellframed images is a limitation of the work; however, there are application domains where this limitation is not overly intrusive. In image databases, for example, the human operator can pre process the image data for objects of interest to be ....

J. Weng, N. Ahuja, and T. S. Huang, "Learning recognition and segmentation using the Cresceptron, " in Proc. International Conference on Computer Vision, pp. 121--128, May 1993.Berlin, Germany.

No context found.

J. Weng, N. Ahuja, and T. Huang, "Learning recognition and segmentation using the cresceptron," in ICCV93, pp. 121--128, 1993.

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