Enhancing edge-based image descriptor models through colour unification

Abstract

The lack of suitable robust appearance models hinders the performance of most image descriptors. Descriptors often rely on pieces of information in images called image features to discriminate the contents of images. Most successful descriptors use gradient images for determining the overall shapes of objects. Consequently, the inferred features are often susceptible to the noise caused by shadows, reflections and inner textures within the object. Significant efforts have been made towards improving the performance of image classifiers, yet generic object detection remains an open problem. In this paper, a method aimed at improving existing appearance models is proposed. The focus is on enhancing the acquired information at fundamental stages to improve the robustness of common statistical learning classifiers, as seen with the work of Holger Winnemoller et al. with human subjects. The selective Gaussian blur filter was applied to several human classification datasets to reduce the amount of ambiguous low frequency noise. Experiments were then conducted to determine whether the unification of similar colours in local image regions could improve the acquired image features. The classification results that were obtained with the processed images were competitive to the results obtained with the original images, however inconclusive for demonstrating the benefits of image smoothing.