Image feature detection and matching for biological object recognition /
Image feature detection and matching are two critical processes for many computer vision tasks. Currently, intensity-based local interest region detectors and local feature-based matching methods are used widely in computer vision applications. But in some applications, such as biological object recognition tasks, within-class changes in pose, lighting, color, and texture can cause considerable variation of local intensity. Consequently, object recognition systems based on intensity-based interest region detectors often fail. This dissertation proposes a new structure-based local interest region detector called principal curvature-based region detector (PCBR) that detects stable watershed regions within the multi-scale principal curvature images. This detector typically detects distinctive patterns distributed evenly on the objects and it shows significant robustness to local intensity perturbation and intra-class variation. Second, this thesis develops a local feature matching algorithm that augments the SIFT descriptor with a global context feature vector containing curvilinear shape information from a much larger neighborhood to resolve ambiguity in matching. Moreover, this thesis further improves the matching method to make it robust to occlusion, clutter, and non-rigid transformation by defining affine-invariant log-polar elliptical context and employing a reinforcement matching scheme. Results show that our new detector and matching algorithms improve recognition accuracy and are well suited for biological object recognition tasks.
School:Oregon State University
School Location:USA - Oregon
Source Type:Master's Thesis
Keywords:optical pattern recognition image processing computer vision
Date of Publication: