Author:

Keywords:

PSI_VISICS

Abstract:

The last decade brought computer vision to a more advanced state. It is the time whenits research results started to influence virtually everybody’s daily life, rather than beingconfined to industrial production lines or other niche applications. From user-interfaceapplications (e.g. the Kinect), over face detection in photos and improved automatedsurveillance, to driver assistance, in all such areas computer vision contributed to theuser-friendliness and safety of consumers’ environment.In this thesis we mainly focus on an important aspect of the algorithms that have broughtthose applications within reach, namely suitable data representations for computervision. The main tasks that directly benefit are classification, detection, segmentation,and image enhancement. Performance, efficiency, and good trade-offs between thoseare the recurrent goals of our research. We consider the representation of vectorialfeatures as linear combinations over pools of samples and this using sparse selectionsor ... all of them. More elaborate field representations are also considered for objectclass modeling and pixel labeling / image segmentation.The contributions include:1. sparse representation-based projections for data dimensionality reduction.2. sparse representations based on Iterative Nearest Neighbors, aimed at closingthe gap between performance (sparse representations - the lasso type) and timeefficiency (nearest neighbors).3. Weighted Collaborative Representations based on the closed-form solutions ofthe Tikhonov regularization.4. different image and feature level representations for Naive Bayes classification.5. the Anchored Neighborhood Regression representation for fast example-basedimage super-resolution.6. starting from sparse and collaborative representations, a training-free classifica-tion framework for textures, materials, and handwritten scores.7. an Elastic Deformation Field Model for representing object classes for detectionpurposes.8. by planar graph representation of images we use the Four Color Theorem fromgraph theory to lower the computational time of the loopy belief propagation insolving Markov Random Field labeling/segmentation problems in early vision.9. efficient representation-driven contributions to applications such as traffic signrecognition and mapping, driver assistance, and extremely fast object detection.