A model-based sampling and sample synthesis method for auto identification in computer vision

Abstract

The need for a large sample size grows exponentially with the dimensionality of the feature space ("curse of dimensionality"), which increases the labor cost during the training procedure and severely restricts the number of the practical applications. While feature selection methods can often alleviate the problems associated with the curse of dimensionality, complex large scale pattern recognition problems may not be amenable to features selection approach due to large intrinsic dimensionality. In such situations, the only effective solution to conquer the complications of the high-dimensional functions is to incorporate knowledge about the data that is correct. How to incorporate the domain knowledge with the specific machine learning system has been widely studied in the pattern classification field. In this paper, we will explore a novel method to synthesize a larger, valid training sample data set based on a smaller set of the key samples that are collected by a model based sampling theory that incorporates the domain knowledge of the computer vision. In addition to reducing the training sample size in the learning procedure, our emphasis is on providing practical advice on how to incorporate domain knowledge to design and simplify a vision based pattern classification model.