A method for noise removal and object detection based on data expansion by multiresolution representation and data compression by principal component analysis
Jean-Baptiste Fasquel, Christophe Stolz, Michel M. Bruynooghe · 2002
This paper presents a method for noisy object detection which is based on the expansion/compression paradigm and combines a multiresolution approach with the principal component analysis (PCA). The multiresolution representation is done by successive Gaussian filterings. The compression of the expanded information is achieved by only keeping the first PCA factorial image. Endly, the object of interest is detected and delineated from the previous factorial image by using a standard valley thresholding technique. The proposed method behaves as a compromise between the various Gaussian filterings by limiting the blurring effect of such filterings and removing most of the noise. The experimental evaluation using synthetic objects has shown the ability of this approach to clean strongly noisy images. For scenes containing several objects of interest, like CTScan images, we first search for regions of interest (ROIs). Then, for each ROI, we locally apply the proposed detection method. Experimental results have shown the potential of the proposed method for the detection of liver tumours from CT-Scan images and for the segmentation of handwritten characters.