A feature selection framework for small sampling data in content-based image retrieval system

摘要

Content-based image retrieval (CBIR) systems have drawn interest from many researchers in recent years. Over the last few years, kernel-based approach has been a popular choice for the implementation of the relevance feedback based CBIR system. This is largely due to its ability to classify patterns with limited sample data. A long flat vector has been a popular choice for the input configuration. The reasons are because it is relatively easy to implement and more importantly, because it preserve the information of identifying the target images via different combination of image features. However, one of the biggest weaknesses of such configuration is the curse of dimensionality. This paper introduces a relevance feedback framework via the use of statistical discriminant analysis method to select only relevant feature for next image retrieval cycle. Hence, minimize the dimensionality of the feature vector. This approach has been tested with four sets of images labelled with different themes. Each set contains 500 images, 50 labelled as positive while the rest are negative. The test showed an improvement from the previous flat input vector configuration when the training samples are relatively small.