State of the art of content-based image classification

Explosive growth of digital technologies has spawned a plethora of images available online. Therefore, content-based image classification has been the subject of many research works in recent years. This paper reviewed some of the most commonly used image classification approaches. Most of the existing approaches used low-level features and intermediate semantic modelling. Image classification using low-level features used colour, texture and shape features directly from the image in combination with learning methods to classify images into several semantic classes (i.e. indoor, outdoor, city, landscape, sunset, forest, etc.). Alternatively, intermediate semantic modelling assigned semantic concepts (i.e. sky, people, grass, etc.) to the image content and the image was classified based on these semantic concepts. Low-level strategies were often used when a small number of categories have to be recognized, as well as when the categories were easily separated. Nevertheless, as the number and ambiguity of the categories increase it was clear that approaches using intermediate semantic concepts were more appropriate.