Image Mosaic Algorithm Based On Feature Point

The image mosaicing technique is a method which is used to match and register agroup of images having partial overlap into a panoramic image. It is commonly appliedin the intelligent transportation, remote monitoring, remote sensing, medicine and otherfields. The effect of image mosaicing is decided by the image registration and imagefusion.Further more, image registration is the core technology in image mosaicing.Nowadays, image registration technology is based on3different methods: graylevel, transform domain and feature. The application of feature registration is mostwidely because of its higher robustness. There are two classic feature registrationalgorithms: SIFT(Scale Invariant Feature Transform) and Harris. The calculation ofHarris corner is simple and the stability is high, but it isn’t scale invariant, so the Harriscorner can’t be used to match the scale invariant images. SIFT is proved to be therobustest feature registration algorithm under the condition of perspective, scale,translation, noise, brightness and rotation. The disadvantages of the algorithm arecomplex operation and low processing speed. In addition, the present registrationalgorithms are effective for the images whose edge are clear.while, for the blur edgeimages are ineffective.SEC (SIFT-Edge-Corner)algorithm is proposed in this paper which is improvedin the feature extration of SIFT. It is present by uniting the advantages of Harris andSIFT. Harris corner point is extracted as the keypoint at the image gaussian edgepyramid, and then the DOG value of keypoints is compared with their adjacentkeypoints, the maximum value is defined as feature point. The experimential resultsshow that the feature extraction efficiency and speed of SEC have been greatlyimproved, and the matching accuracy has been improved to some extent.In order to solve the problem that the registration of edge fuzzy images, theextreme gray offset feature is proposed. Firstly, the feature point is defined as the piexlwith the most dramatic offset from the local mean value. After that, the extreme offsetfeature is located in sub-pixel level by derived function, while the function curve isfitted by the least square. Finally, the registration is realized by comparing the distanceamong those32d feature vectors which are built by NMI. This feature is easy to extractand calculate, the results of experiments show that the algorithm can effectively match deeply fuzzy images, and it is invariant in rotation, translation, perspective andbrightness at the same time.Phase correlation algorithm is used to sort images with simple translation in somestudies.The phase correlation algorithm under the polar phase correlation andlogarithmic polar coordinates is introduced into the sorting algorithm in this article so asto enhance the applicability of the algorithm. At last, the gradually fade image fusionalgorithm is selected to complete the image stitching.