AbstractsEngineering

Local feature detection and local feature description have been frequently utilised as an integral system to accomplish diverse computer vision applications due to its invariance capabilities against image distortions caused by geometric and photometric transformations. The applications include image retrieval, image registration and object recognition. In computer vision, however, local feature detection and description should be treated as independent processes since the former identifies repeatable keypoints in images while the latter analyses local image patches around detected keypoints to represent those keypoints. Following the principle of separable processes, this research concentrates on local feature description and suggests further improvements. A variety of local feature descriptors has been devised based on statistical analysis in spatial domain such as creating a histogram of image gradients. To address the overdevelopment of these approaches, a descriptor that relies on spectral analysis in frequency domain is proposed. The proposed descriptor, which is coined as Radial Fourier Analysis (RFA) descriptor, transforms the image gradients of local image patches to frequency domain, then its frequency response is decomposed and analysed to provide a robust and distinctive representation for keypoints. Concurrent with the use of spectral analysis in the frequency domain, a keypoint orientation estimator is designed in this domain to enhance the rotational invariance of RFA descriptor. The keypoint orientation estimator employs the starting point normalisation of Fourier coefficients to deduce rotating angles that ensure the local image patches of keypoint correspondences are aligned at the same orientation. As such, the performance of RFA descriptor is further enhanced with this accurate estimation of keypoint orientation. Through comprehensive experiments and comparisons, RFA descriptor had demonstrated its outstanding performances against various image distortions. In particular, the descriptor gives reliable performances in dealing with images that are degraded by blur, rotation, illumination and JPG compression changes. It also indicates that the proposed RFA descriptor and keypoint orientation estimator combine perfectly to yield a powerful local feature descriptor. All these show that spectral analysis serves well as an alternative approach for statistical analysis of local feature description and keypoint orientation estimation.