Multi-directional Multi-resolution Transforms for Zoom-Endoscopy Image Classification

Abstract

Summary. In this paper, we evaluate the discriminative power of image features, extracted from subbands of the Gabor Wavelet Transform and the Dual-Tree Complex Wavelet Transform for the classification of zoom-endoscopy images. Further, we incorporate color channel information into the classification process and show, that this leads to superior classification results, compared to luminance-channel based image processing.

Computer-aided colorectal tumor classification in NBI endoscopy using local features

“…Kwitt et al (2010) introduced a generative model that involves prior distributions as well as posteriors, and employed a two-layered cascade-type classifier that achieved 96.65% for two classes and 93.46% for three classes. Other work from this group includes texture analysis with wavelet transforms (Häfner et al, 2009f;Kwitt and Uhl, 2007a), Gabor wavelets (Kwitt and Uhl, 2007b), histograms (Häfner et al, 2006), and others. In our previous work (Takemura et al, 2010), we have used shape analysis of extracted pits, such as area, perimeter, major and minor axes of a fit ellipse, diameter, and circularity.…”

Survey of Computer Vision and Machine Learning in Gastrointestinal Endoscopy

“…In [13], authors evaluated the discriminative power of image features extracted from sub-bands of the Gabor and the Dual-Tree Complex Wavelet Transform for the classification of zoom-endoscopy images. Further they also incorporated colour channel information and show, that this leads to superior classification results, compared to luminance-only based processing.…”

Computer-assisted pit-pattern classification in different wavelet domains for supporting dignity assessment of colonic polyps

“…In (Kwitt et al, 2008) the feature extraction step is based on the Gabor Wavelet Transform (Manjunath and Ma, 1996) and the Dual-Tree Complex Wavelet Transform (DT-CWT) (Kingsbury, 1998), which both provide approximate shift-invariance and a directionally-selective frequency partitioning. Color-information from the RGB color-channels is incorporated by feature vector concatenation or by using a parallel multi-classifier.…”

“…Although it is true that we might lose information by neglecting color-channel dependencies, we will see in Section 4 that concatenation is a quite good combining scheme for our problem. Another representative of non-integrative approaches is presented in (Kwitt et al, 2008), where a parallel multi-classifier is used to incorporate color information. However, the results presented there show that concatenation is superior to the parallel multi-classifier in almost all cases.…”

“…For simplicity we decided to use a 1-NN classifier for all experiments. On the basis of different feature sets, other work (Häfner et al, 2007a,b;Kwitt and Uhl, 2007;Kwitt et al, 2008) has shown that 1-NN lead to very good results. Since the scope of this paper is on the feature computation side and not on testing classifiers, we limit our discussions to the 1-NN case.…”