Ensemble of steerable local neighbourhood grey-level information for binarization

• Published in: Pattern recognition letters Vol. 98; pp. 8 - 15
• Main Authors: Kasmin, F., Abdullah, A., Prabuwono, A.S.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 15.10.2017; Elsevier Science Ltd
• Subjects: Background noise; Binarization; Binary system; Ensemble; Filters; Grey level; Image classification; Image coding; Image manipulation; Image processing; Image processing systems; Local neighbourhood; Noise reduction; Pattern recognition systems; Pixels; Retina; Retinal images; Steerable; Support vector machines; Vision systems; Steerable; Local neighbourhood; Binarization; Grey level; Ensemble
• ISSN: 0167-8655; 1872-7344
• DOI: 10.1016/j.patrec.2017.07.014

•We combine a set of steerable SVM filters for binarization.•We construct an ensemble of SVMs and evaluate its performance on four datasets.•The performance of our ensemble is comparable to that of the existing approaches. Steerable filters are very useful in vision and image processing because the response of the filters at various orientations can be examined and manipulated, which is useful for texture analysis and image denoising. In supervised binarization, a set of grey-level values can be used to represent a particular pixel and to determine whether it belongs to the foreground or background. However, extensive noise may be produced in the resultant images if the chosen grey-level values are insufficient to describe the pixel. This may occur when some of the pixels are wrongly classified. To overcome this problem, the advantages of steerable filters are employed in proposed steerable local neighbourhood (SLN) of grey-level information methods to characterise pixels in images. The proposed methods use a support vector machine to classify each pixel using SLN grey-level information. Sets of normalised intensities of grey-level values are generated according to the orientations at 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315°. These sets of feature vectors contain the vectors of the local neighbourhood of grey-level information of each pixel. Document and retinal images are used to train and test the accuracy of the proposed classifier. On the basis of the results obtained for the training images, weights are applied to every SLN. Then, the ensemble of steerable local neighbourhood methods, namely the weighted addition rule and the weighted product rule, are used to combine all the SLNs’ grey-level information. The results of the proposed methods are promising and clearly show a significant improvement in terms of accuracy compared to that of other methods.