Pattern Recognition

Abstract

Pattern recognition (PR) concerns the description or classification (recognition) of measurements. PR capability is often a prerequisite for intelligent behavior. PR is not one technique, but rather a broad body of often loosely related knowledge and techniques. PR may be characterized as an information reduction, information mapping, or information labeling process. Historically, the two major approaches to pattern recognition are statistical (or decision theoretic), hereafter denoted StatPR, and syntactic (or structural), hereafter denoted SyntPR. The technology of artificial neural networks has provided another alternative, neural pattern recognition, hereafter denoted NeurPR. NeurPR is especially well suited for “black box” implementation of PR algorithms. As no single technology is always the optimal solution for a given PR problem, all three are often considered in the quest for a solution.

Assessment of Driver Stress from Physiological Signals collected under Real-Time Semi-Urban Driving Scenarios

“…The stress-classification neural network consists of an input layer, which takes in an input feature vector representing the affective state of drivers, a hidden layer to model the nonlinearities in data and an output layer to predict the target affective state class 34 . Such networks employ a connectionist approach to compute the interconnection weights and bias parameters which produce the most optimal configuration 35 .…”

Opportunity and challenge on face recognition: generic-to-specific feature representation and recognition strategy

“…The form of the decision boundary used can be either linear or non-linear. [15] [11] [13] In the statistical approach, the size of the feature for each object is described as having the same dimension and the relationship between feature are not considered so that the results of the introduction are difficult to explain through human logic and system performance is highly dependent on the distribution of training results data. The weakness in the statistical approach is the ability to overcome the emerging variance of an object arising from differences in illumination, differences in acquisition devices, differences in data acquisition orientation, natural changes in objects over time.…”

“…Through a structural approach, the size of the representation for each object can be different and allows for combinatorial combustion in the representation size. It will require a large measure of training data and computational resources [11].The representation of the object in the structural / syntactic approach can be a string, graph, tree, PDL (Picture Description Language) where the inference process can use grammar, finite-state-automata approach, string matching, Hopfield network, matching graph / tree [13] [15].…”

Focusing Keywords to Automatically Extracted Image Segments Using Self-Organising Maps

“…In the first step ISODATA variant of K-means algorithm [40] with a K value 15 is used to compute an oversegmentation based on the colour coordinates of the pixels. This step typically results in a few thousand separate segments.…”