Abstract
Noise is unwanted signal that causes a major problem for the task of image classification and retrieval. However, this paper reports that adding noise to texture at certain levels can improve classification performance without training data. The proposed method was tested with images of different texture categories degraded with various noise types: Gaussian (additive), salt-and-pepper (impulsive), and speckle (multiplicative). Experimental results suggest that the inclusion of noise can be useful for extracting texture features for image retrieval.