Content-based Image Classification via Visual Learning

Abstract

As the information processing ability of computers improves, real-world images are increasingly used in various data mining applications. Thus, flexible and accurate image recognition methods are strongly needed. However, real-world images generally contain a wide variety of objects which have complex features (e.g., shapes and textures). Therefore, the accurate recognition of real-world objects is difficult because of three main problems. Firstly, although an image is generally given as a set of pixels, pixels alone are insufficient for the description and recognition of complex objects. Thus, we must construct more discriminative features from pixels. Secondly, finding useful features to describe complex objects is problematic because appropriate features are dependent on the objects to be recognized. Thirdly, real-world images often contain considerable amounts of noise, which can make accurate recognition quite difficult. Because of these problems, the recognition performance of current recognition systems is far from adequate compared with human visual ability. In order to solve these problems and facilitate the acquisition of a level of recognition ability comparable to that of human visual systems, one effective method consists of introducing learning schemes into image understanding frameworks. Based on this idea, visual learning has been proposed (Krawiec & Bhanu, 2003). Visual learning is a learning framework which can autonomously acquire the knowledge needed to recognize images using machine learning frameworks. In visual learning, given images are statistically or logically analyzed and recognition models are constructed in order to recognize unknown images correctly for given recognition tasks. Visual learning attempts to emulate the ability of human beings to acquire excellent visual recognition ability through observing various objects and identifying several features by which to discriminate them. The key to the development of an efficient visual learning model resides in features and learning models. Image data contain various types of informative features such as color, texture, contour, edge, spatial frequency, and so on. However, these features are not explicitly specified in input image data. Therefore, feature construction is needed. Feature construction is the process of constructing higher-level features by integrating multiple lower-level (primitive) features. Appropriate feature construction will greatly contribute to recognition performance. In addition, since useful features depend on the given image data, O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg