Image classification using generalized multiscale RBF networks and discrete cosine transform

The use of the multiscale generalized radial basis function (MSRBF) network for image feature extraction is proposed for the first time. The MSRBF network holds a simple but flexible structure capable to modelling complex systems. However MSRBF is originally designed to identify observational-type input-output systems. We aim to use this efficient network to get to concise but accurate models of digital images thanks to: a) the use of multiple scales in the RBF kernel width, and b) the adoption of the forward regression orthogonal least squares (FROLS) algorithm to refine the model structure selection. Thereafter the new tailored model is excited to produce output signals aimed at be compressed by the discrete cosine transform (DCT), adopted in this work to compact signals’ energy into a few coefficients. To recognise images as MSRBF networks, a mathematical modelling was done by considering the first ones as multiple-input single-output systems. Based on the new methodology a novel computer aided diagnosis (CAD) system for cancer detection in X-ray mammograms was designed. Classification results show that the new CAD method helped reach a competitive diagnostic accuracy of 93.5%. It was similarly found that the MSRBF network is able to construct tailored and precise image models.