|
Lee, Te-Won ICA Mixture Models for Image Processing The ICA mixture model was proposed for unsupervised classification and automatic context switching in blind source separation (Lee, 1999 NIPS). Here, the algorithm is applied to image processing problems such as unsupervised image classification, segmentation, de-noising, and feature extraction. We demonstrate that the algorithm is effective in classifying complex image textures such as trees and rocks in natural scenes. For de-noising and filling in missing pixels in images, we show that the algorithm codes the images more efficiently due to the flexibility of the mixture model and thus reduces noise better than traditional techniques. The classes of extracted image codes can be used in conventional pattern recognition systems for improved performance.
|