Research On Manifold Perception-Based Image Distance Metric

Human being’s activities are heavily depended on rapid and accurate identification ofobjects in field of vision. Despite it is easy for human being’s brain, image recognition isdifficult for computer and is widely acknowledged as a challenge in computer vision.Measuring the distance or similarity between images is a fundamental problem in imagerecognition. A reasonable image distance metric must be able to finely discriminatebetween different objects while tolerant to object variations such as rotation, scaling,non-rigid deformations and so on. Recently manifolds are thought to be fundamental forhuman being’s vision perception. Retinal images of one object according to continuousvariability of rotation, scaling and so on lie on a low-dimensional image manifoldembedded in high-dimensional obsering space. How to use biological clues in imagedistance metric is of great significance.In this dissertation, the process of human being’s manifold perception is simulated, severalimage distance metrics based on manifold perception are proposed and their applications inimage recognition are addressed. The main research works are described below.Firstly, a high-order approximated manifold distance (HMD) based on manifoldlearning and high-order manifold approximation was proposed for image distancemeasurement. The intrinsic variables of nonlinear image manifold are learned by a specialmanifold learning algorithm–Maximum Variance Unfolding (MVU), and the nonlinearimage manifold is meshed into grid along the directions of intrinsic variables. Thehigh-orde partial derivatives of image manifold with respect to intrinsic variables areapproximated by finite difference. The nonlinear image manifold is approximated by thecurve surface based on high-order Taylor expansion of image manifold with respect tointrinsic variables. HMD is defined as the minimum distance between the query image andthe approximated curved surface of image manifold in image space. HMD can be directlyutilized in distance-based classifiers for image recognition. The experimental results offace and handwritten digit recognition demonstrate that HMD not only attain larger rangeof invariance to image transforms but also achieves higher recognition accuracy andstability than several state-of-the-art image distance metrics, such as Euclidean Distance,Tangent Distance, Image Euclidean Distance (IMED) and so on.Secondly, a Gabor feature-based approximated manifold distance (GFMD) was proposed which measures image similarity in Gabor feature space. Because theredundancy of original image influences the accuracy of image distance measurement, thetwo-dimensional Gabor wavelets with different scales and orientations are applied toproduce salient local and discriminating features of original images and augumentd Gaborfeature vectors are derived. The intrinsic variables of Gabor feature manifold are revealedby MVU, and Gabor feature manifold is approximated by the curve surface based onhigh-order Taylor expansion of feature manifold with respect to intrinsic variables. GFMDis defined as the minimum distance between the Gabor feature of query image andapproximated curved surface of Gabor feature manifold in feature space. GFMD can bedirectly utilized in distance-based classifiers for image recognition. The experimentalresults of face and handwritten digit recognition demonstrate that GFMD achieves higherrecognition accuracy and stability than HMD and other state-of-the-art distance metricscombined with Gabor features. However, GFMD has higher computational complexitythan these compared distance metrics.Thirdly, a complex object recognition model based on visual attention model andsubmanifold distance (SMD) is developed. Objecet-accumulated visual attention model isselected for object detection in clutter, which iteratively extends current attentional regionin order to extract the whole region of target. The image patches corresponding to currentattentional regions constitute multiple submanifolds of target. In learning stage thesubmanifolds are acquired by clustering the image patches. The intrinsic variables of eachsubmanifold are learned by MVU, and each submanifold is approximated by curve surfacebased on high-order Taylor expansion of submanifold with respect to intrinsic variables.SMD is defined by the minimum distance between queried image patch and theapproximated curved surface of submanifold. In recognition stage SMD between imagepatch of current attentional region and target’s submanifolds is computed. The extendingstrategy of current attentional region is controlled by SMD. The experimental results ofcomplex object recognition in remote sensing images illustrated that our proposed modelcan recognize complex objects in clutter accurately, robustly and quickly.