A Scale Invariant Detector Based on Local Energy Model for Matching Images

Finding correspondent feature points represents a challenge for many decades and has involved a lot of preoccupation in computer vision. In this paper we introduce a new method for matching images. Our detection algorithm is based on the local energy model, a concept that emulates human vision system. For true scale invariance we extend this detector using automatic scale selection principle. Thus, at every scale level we identify points where Fourier components of the image are maximally in phase and then we extract only feature points that maximize a normalized derivatives function through scale space. To find correspondent points a new method based on the Normalized Sum of Squared Differences (NSSD) is introduced. NSSD is a classical matching measure but is limited to only the small baseline case. Our descriptor is adapted to characteristic scale and also is rotation invariant. Finally, experimental results demonstrate that our algorithm is reliable for significant modification of scale, rotation and variation of image illumination.