Neural network for visual search classification

Abstract

-Visual search describes the process of how the eyes move in a visual field in order to acquire a target. Visual search needs to be quantified to improve future search strategies. This paper describes a hybrid neural network used to perform v isual search classification. The neural network consists of a Learning vector quantization network (LVQ) and a single layer perceptron. The objective of this neural network is to classify the various human visual search patterns into predetermined classes. The classes signify the different search strategies used by individuals to scan the same target pattern. The input search patterns are quantified with respect to an ideal search pattern, determined by the user. A supervised learning rule, Learning vector quantization1 (lvq1) is used to train the network. I. INTRODUCTIONThe present project relates to a LVQ network for pattern classification, which is able to vary it's response signal by learning to separate and to identify a correct class of the input signal from repeated presentations of an input pattern signal. Vector quantization traditionally is a technique used for compression of speech and image data [7]. The aim of the network is to classify visual search. During free viewing where there is no objective, eye position can be modeled as a Markov chain with stable probabilities to describe the position of the fovea. During timed visual search, eye movements tend to be more structured [6]. People follow a very generalized subject dependent search pattern at first, and then follow a very generalized search pattern not dependent on subject but based on the task at hand [7]. There are varieties of schemes employed, such as horizontal step-down function, processing the display in a column format, spiral in or out etc. In this research, the hybrid neural network consists of a LVQ network that uses a supervised compet itive learning scheme and a single layer perceptron that uses the leastmean-square (LMS) learning algorithm. The network is simulated in Matlab using the lvq function.The goal is to classify, compare and quantify visual search patterns. The work described in this paper is the first step towards reaching that goal. II. METHODOLOGYThe architecture of a LVQ network is very similar to a self-organizing memory. In both networks each neuron excites itself and inhibits all the other neurons. The neuron (i.e. ith processing element) whose weight vector is closest to the input (i.e. winning neuron) has the largest net input and its output is set to one. All other outputs are set to zero. The network modifies the weight of the winning neuron as a function of the error (i.e. the difference between the input vector P(n) and the weight vector Wi(n)) and the learning rate á.Wi is the weight matrix if the i th processing element. Wi (n+1) is new modified weight matrix. Wi (n) is the current unmodified weight matrix. P(n) is the input vector. á is the learning rate.The problems that a self-organizing memory has are the tradeoff between fast learning and stability and t...