Coarse adaptive color image segmentation for visual object classification

This paper deals with perceptually inspired image segmentation for the purpose of generic image classification or object detection. Indeed, in our algorithm we will try to stay true to human perception and more specifically Gestalt Theory. Input images are processed in a three-step framework: a pre-processing step where the image is filtered and where perceptually similar colors are grouped as per color constancy law, a clustering step where we also determine an optimal number of quantized colors and a post processing step where we add spatial information and merge smaller regions as per ldquogood continuationrdquo and proximity laws. Another major feature of our algorithm is that it adapts to image dynamics and doesnpsilat require image-specific parameter tuning. Application on a 10,000 image dataset shows the algorithm succeeds in producing large coarse regions that can be used for feature extraction.