Learning an Image Lexicon

In this work, we propose a methodology to train a classifier in learning the mappings between images and their corresponding class labels in words. We describe recognition as a process of annotating image regions with words. We consider this task as learning a lexicon for a short image vocabulary that will translate image regions to words. We consider images wherein only one conceptual object, such as ´building´ is apparent in the entire image. Objects are manually cropped from other images. Then, we extract color descriptor in the L*a*b* color space, and texture features based on information obtained from the windowed image second moment matrix that provide us with three pieces of information about each pixel such as anisotropy, polarity, and contrast. A feedforward neural network with improved generalization is trained to learn the mapping between descriptors of the images and class labels. Our model is developed based on a dataset of 2800 images. Experimental results are illustrated to observe the strength and weaknesses of the model. We provide a comparison table that shows the supremacy of our methodology over other model