A Study On Orientation Estimating Of The Descriptor Based On Intensity Order

In recent years, represented by SIFT local descriptor has achieved great success, promoting the related research progress in the field of computer vision. At present, the local descriptor matching has become a major paradigm of wide baseline matching, object recognition, image classification and image retrieval, etc. In many applications, a major paradigm is driving these applications and in turn it also leads to a further study of local descriptor.Local descriptor structure usually includes several links or steps, the specific operation of each link will affect the descriptor’s final performance. Rotation invariance is to describe the basic requirements in order to ensure that descriptor rotation invariance. Mostly, describing the step includes a calculation of orientation of the intermediate links. But the latest statistics show that the present method of calculating orientation has almost a third of its error more than20degrees, which greatly reduces the performance of the descriptor. So a stronger robustness calculation method of orientation will play an important role in promoting.Aiming at this problem, this article explores a new kind of calculation method for orientation, and a good result has been achieved, specific research contents and research results are as follows:1. Orientation is proposed based on intensity order of construction method. Characteristics is the basic idea to support regional grouping pixels according to intensity order, and then computing group within the pixels of geometry, statistics, and on this basics we calculate orientation. This thesis gives two pixel grouping methods, which are based on average intensity level and the method based on the average number of pixels. It also gives two kinds of different geometrical statistic calculation methods for the orientation. Respectively is based on the geometric center of gravity point and based on ellipse fitting method. Finally, in view of the different scenarios, the image rotation, scale changes, illumination, image blur, image compression, which has carried on the performance test under complicated illumination change. In majority of cases, its results are better than previous methods.Different from gradient method, this paper is based on the intensity order and its main advantages are as follows:First, the image intensity itself is smaller than gradient information when they are affected by noise; Second, based on the gradient method of intensity linear transformation is variable as long as intensity order unchanged; Third, in this paper, the method of orientation is continuous, but the previous methods are discrete data by interpolation; Fourth, in term of the computational complexity, this paper adopted the lookup method, whose computation complexity is low.2. Orientation is given on the properties of descriptor impact analysis. Orientation of calculation is primarily for the descriptor. This article puts forward orientation of application in the classical description and standardized performance tests, the results shows that the descriptor properties generally have a larger ascension. It also shows that the proposed calculation method of orientation is a very practical method.