Difference of Gaussian (DoG) is an effective feature which can discriminate between presence and absence of a blob object in an image. A DoG scale-space as shown on the left is a collection of images usually forming an image pyramid. We defined a kernel function between two DoG scale-spaces by weighted summation of circular correlation between DoG response images at the same scale. Multiple kernel learning (MKL) refers to learning these weights. We proposed an algorithm where we simultaneously learn MKL by sparse linear prediction and a support vector machine to classify images having blob object from those not having any. This method was applied to some industrial applications.

Related Publications

• S. Nilufar, N. Ray and H. Zhang, “Object detection with DoG scale-space: A multiple kernel learning approach,” IEEE Transactions on Image Processing, vol.21, no.8, pp.3744-3756, 2012.
• S. Nilufar, N. Ray, H. Zhang, “Wavelet subband-based steam detection by multiple kernel learning,″ IEEE ICIP 2012.
• S. Nilufar, N. Ray, M.K.I. Molla, K. Hirose, “Spectrogram based features selection using multiple kernel learning for speech/music discrimination,” IEEE ICASSP 2012.