Scale space feature selection with multiple kernel learning and its application to oil sand image analysis

Scale-space representation for an image is a significant way to generate features for object detection/classification. The size of the object we are looking for as well as its texture contents are related to the multi-scale representations. However, any scale-space based features face the inevitable issues of high dimentionality and scale selection. Scale-space analysis of image provides a set of extremely high dimensional features at each scale- the number of pixels in a filtered output image is the feature dimensionality at that scale. Moreover, considering all the output images at various scales, the dimensionality of the feature set is staggeringly high. Selection of features from this high dimensional space is daunting. In addition, the scale selection process is still ad-hoc, while applying scale-space based features for object detection/classification. In this research these two issues are resolved by designing a suitable kernel function on the scale space based features and applying multiple kernel learning (MKL) approach for sparse selection of scales.;A novel shift invariant kernel function for scale space based features is designed here. Also a novel framework for multiple kernel learning is proposed that utilizes a 1-norm support vector machine (SVM) in the MKL optimization problem for sparse selection and weighting of scales from scale-space. The optimized data-dependent kernel accommodates only a few scales that are most discriminatory according to the large margin principle. With a 2-norm SVM this learned kernel is applied to the classification problem.;In this thesis we have applied the proposed classification method for oil sand image analysis. Automatic analysis of oil sand video images is non-trivial due to the presence of dirt and fine materials. In addition, changeable weather and lighting condition make the video quality worse. Two challenging problems in oil sand mining which are detection of large lump and steam from videos are investigated here. Difference of Gaussian (DoG) and wavelet scale space are applied for these two different detection problems, respectively. Our method yields encouraging results on these difficult-to-process video images and compares favourably against other existing methods.