Art Object Recognition by Image Processing based on Color and Hidden Markov models (2009)
Nowadays, most of the daily information we use is digital, whether in the arts or at our work. However, images are often stored in huge, unclassified databases where it is extremely difficult for people to search for images with specific or desirable features. For example, an artist or an art teacher interested in painting may want to retrieve images with a specific color hue or a particular texture. Therefore, it is necessary to provide users with a searching tool that helps them retrieve images according to specific characteristics.
This work proposes a searching tool based on image colors. The approach involves two steps:
−Color features extracted using color histograms: It computes the statistical distribution of the colors in an image. In order to take into account both color distribution and spatial information, weighted histograms are used. Weights (usually local laplacian or entropy) are needed to compute the probability of a color in its neighbourhood. Weighted histograms are an efficient way to retrieve similar images according to their color information.
−A training model based on HMM (Hidden Markov Models): Markov models constitute the most successful approach developed for modeling the statistical variations in temporal pattern recognition applications. They have been proved very efficient in speech recognition, and are nowadays implemented in recent operating systems (Vista, MacOS, Linux). A Markov process is a system which can be described at any time as a set of N distinct states. In the case of color images, Markov states correspond to color packets. Indeed, a color image usually contains only hundreds or a few thousands different colors among a theoretical choice of 16,777,216 (typical RGB color model). This means many colors are missing, and this information is used to compute packets of continuous colors. HMM-based systems must first be trained and validated with data samples in order to acquire their probabilities of transitions between states. Once trained and validated, the system can be used to classify learned or new patterns.
This method has been tested on the image database COIL (Columbia Object Image Library) which contains 7,200 images of 100 different objects, each object being viewed under 72 different angles. The optimal results were obtained by using one third of the images of each object for the training phase, and the rest for validation. More specifically, every image is previously indexed by its color histogram information from which the set of color packets is used as Markov states. The rate recognition success on COIL reaches 90%, meaning that 90% of the images are correctly assigned to their corresponding object. In order to improve the recognition rate, we plan to extract other visual image features like shapes or textures as well as to be able to process databases with higher number of images.
Eric Hervet, Ph.D. and Mustapha Kardouchi, Ph.D. are part of the Département d’informatique at the Université de Moncton.