Object localization is currently an active area of research in computer vision. The object localization task is to identify all locations of an object class within an image by drawing a bounding box around objects that are instances of that class. Object locations are typically found by computing a classification score over a small window at multiple locations in the image, based on some chosen criteria, and choosing the highest scoring windows as the object bounding-boxes. Localization methods vary widely, but there is a growing trend towards methods that are able to make localization more accurate and efficient through the use of context. In this thesis, I investigate whether contextual relationships between related objects can be leveraged to improve localization efficiency through a reduction in the number of windows considered for each localization task. I implement a context-driven localization model and evaluate it against two models that do not use context between objects for comparison. My model constrains the search spaces for the target object location and window size. I show that context-driven methods substantially reduce the mean number of windows necessary for localizing a target object versus the two models not using context. The results presented here suggest that contextual relationships between objects in an image can be leveraged to significantly improve localization efficiency by reducing the number of windows required to find the target object.

This thesis describes an approach for accomplishing a pattern recognition task using conceptual graph theory and neural networks (NNs). The set of patterns to be recognized are the capital letters of six different fonts of the English alphabet, plus two shifted and six rotated versions of each. The letters are represented to the neural network on a 16x16 input grid (256 "sensor lines"). A standard classification encoding for such patterns is to use a 26-bit vector (26 lines at the NN's output), one bit corresponding to each letter. Experiments with such an encoding yielded results with poor generalization capability. A new encoding scheme was developed, based on the conceptual graph formalism. This entailed designing a set of concepts and a set of associated relations appropriate to the upper case letters of the English alphabet. From these, the following were developed: a conceptual graph representation for each letter, a connection matrix for each, and finally, a C-vector and an R-vector representation for each. The latter were used as the output encoding (21 bits) of the NN pattern recognizer. A feed-forward neural network with 256 inputs, 21 outputs, and 2 hidden layers was trained using the back-propagation- of-error algorithm. Results were significantly better than with the more standard. encoding. Generalization on fonts improved from 74% to 96%, generalization on rotations improved from 83% to 94%, and finally, generalization on shifts improved from 2% to 93%.

Note:

No description available.

No description available.

No description available.

This paper presents an algorithm for a dollar bill recognition system. Although this thesis describes it in detail for the specific application of designing a dollar bill recognition system, the algorithm is quite general and can be applied to a variety of pattern recognition problems. The scheme operates on the image of a corner of the bill. Hough transform is used to find the edges and the corner point in the image. If there is any skew in the edges, it is corrected and a 256 x 256 pixel image is obtained. This image is then compressed to an 8 x 8 matrix, and features are extracted from a two dimensional Walsh Transform of this matrix. The process of feature selection is based upon the standard deviations of the Walsh coefficients. These features are then used by a Sequential Classifier for classifying the bill. / Master of Science

This work describes the design of an intelligent recognition system used to distinguish noise signatures of five different acoustic sources. The system uses pattern recognition techniques to identify the information obtained from a single microphone. A training phase is used in which the system learns to distinguish the sources and automatically selects features for optimal performance. Results were obtained by training the system to distinguish jet planes, propeller planes, a helicopter, train, and wind turbine from one another, then presenting similar sources to the system and recording the number of errors. These results indicate the system can successfully identify the trained sources based on acoustic information. Classification errors highlight the impact of the training sources on the system's ability to recognize different sources. / Master of Science

A system for the extraction and storage of line and region data from digitized engineering line drawings, first proposed by Watson et.al.[3] and further developed by Bixler et.al.[4], is completed. As a means for the automatic analysis of picture content, a model based recognizer for line patterns is developed. The pattern matcher uses a simple scheme to decompose a line drawing into basic parts: strokes and junctions, and then finds graph isomorphisms between known line pattern models stored in a database and portions of the image line data. Hu's moment invariants [16] are used to match simple shapes and prune the search space. Information about the connectivity of patterns matched in the image is retained, allowing higher level analysis of image content. A second method for calculating a moment signature from line data is presented. This method makes use of a spline approximation of the line data and Legendre polynomials. Some methods for recognizing incomplete line patterns and partially occluded curves are also discussed, and some experiments are performed. / M.S.

Two methods of designing a point classifier are discussed in this paper, one is a binary decision tree classifier based on the Fisher's linear discriminant function as a decision rule at each nonterminal node, and the other is a contextual classifier which gives each pixel the highest probability label given some substantially sized context including the pixel. Experiments were performed both on a simulated image and real images to illustrate the improvement of the classification accuracy over the conventional single-stage Bayes classifier under Gaussian distribution assumption. / Master of Science