Image Matching And Recognition Based On Local Features

Image matching is one of most important research topics in the field of computer visionand pattern recognition. In recent years, image matching has been developing very fast andhas been applied in the fields of medical image analysis, biometrics, and robotic control, etc.The objective of image matching is to find correspondences between images. Among vari-ous image matching methods, the methods based on local features has become an emphasisin related research fields due to their excellent invariance properties with respect to variousimage transformations.Local features-based image matching approaches are invariant to image variancecaused by changes in imaging conditions,like scale change, viewpoint change and illumi-nation change. In addition, the local features are distinctive and are able to be matchedaccurately even in large datasets. Local features-based image matching approaches consistin three steps: feature detection, feature description and feature matching. Local featuredetection is a key issue in image matching because it sets limits on the potential invarianceproperties and matching performance.The now-standard DoG, Harris-Laplace and Hessian-Laplace detectors embed the ba-sic detector in the Gaussian scale-space and utilize the scale selection technique to achieveinvariance to scale transfor-mation. Though the Gaussian function is the unique continuouskernel that satisfies the causality property, the discretized Gaussian kernel used in practiceno longer possesses the causality property. In addition, the Gaussian derivative kernels arenot well matched to the local structures in images. Finally, the computational complexityincreases as the scale gets larger, which makes the keypoint detectors not suitable for ap-plications that have real-time constraints. In this paper, we study local feature detection indepth and accomplish the following works:(1) we investigated the b-spline scale-space in detail (e.g., b-spline derivatives are betteradapted to the image structures) and demonstrate the advantage of b-splines compared toGaussian in scale selection. (2) We proposed an efficient discrete convolution algorithm for implementing b-splinescale-space by extending the integral image technique.(3) We took advantage of the determinant of the Hessian matrix based on b-splinescale-space and proposed a scale-invariant feature detector which could detect both bloband junction structures.(4) We evaluated the performance of exiting feature detectors and the proposed oneon the Oxford Affine Covariant Features dataset (repeatability and matching performance).The results demonstrate that the proposed detector is efficient and outperforms DoG, SURFand Hessian/Harris-Laplace detectors in terms of invariance and distinctiveness.