| In recent years, target recognition motivated by the neuroscience of the visual cortex has become a hot machine learning issue in the area of pattern recognition and image processing. A good similarity measure, which can unify the selectivity of different targets in an image and the invariance of one target in distinct images, is needed in this kind of problems, especially to the target recognition in complex environment. Consequently, exploring some novel efficient algorithms is still a study hotspot in the research area.The paper overviews the research actualities of image matching algorithm, especially the studies on similarity metric, and analyzes the constructed process of derived kernel model and some problems encountered by it. The basic idea of derived kernel model is introducing the responding principle of primate visual cortex into the machine learning research and using some cost function to measure the similarity of images'internal features. According to the structure of derived kernel algorithm, the paper proposes a novel algorithm called derived kernel algorithm based on log-polar transformation (LPT). Subsequently, the basic ideas, mathematic model and the selection of templates are elaborated and the invariance of this new method is also proved. Simulation experiments on face recognition have demonstrated the good performance of this algorithm in handling some complex transformations, like rotation, scale and so on. Finally, the defects of this new method are analyzed and some future research directions are discussed.Derived kernel algorithm, especially its translation invariance, has provided some foundation for the researches in this area. Based on this, a novel derived kernel algorithm based on log-polar transformation has been proposed in this paper, which has a good performance on target recognition in complex images with rotation and scale. The rotation and scale invariance of this new approach has been demonstrated both in theory and in practice. |
PDF Full Text Request