Efficient image selection for concept learning

In semantic-based image classification, learning concepts from features is an ongoing challenge for researchers and practitioners in different communities such as pattern recognition, machine learning and image analysis, among others. Concepts are used to add knowledge to the image descriptions linking high- and low-level numerical interpretation of the image content. Augmented descriptions are useful to perform more ‘intelligent’ processing on large-scale image databases. The semantic component casts the classification into the supervised or learning-from-examples paradigm, in which the classifier obtains knowledge by generalising specific facts presented in a number of design samples (or training patterns). Consequently, selection of suitable samples becomes a critical design step. The introduced framework exploits the capability of support vector classifiers to learn from relatively small number of patterns. Classifiers make decisions based on low-level descriptions containing only some image content information (e.g. colour, texture, shape). Therefore there is a clear drawback in collecting image samples by just using random visual observation and ignoring any low-level feature similarity. Moreover, this sort of approach set-up could lead to sub-optimal training data sets. The presented framework uses unsupervised learning to organise images based on low-level similarity, in effort to assist a professional annotator in picking positive and negative samples for a given concept. Active learning to refine the classifier model follows this initial design step. The framework shows promising results as an efficient approach in selecting design samples for semantic image description and classification.