Study On Local Descriptor

Along with the fast growth of the ways and amount of humans'image gaining, some basic researches in digital image processing became more and more important. Among them, local descriptor is one of the most basic and most difficult fields in computer vision and pattern recognition. Generally, local feature is an image pattern which differs from its immediate neighborhood. They are distinctive, do not require segmentation and robust to occlusion, overlap, geometric transformation and illumination change. Because of these characteristics, local feature become the most important technology in feature extraction.Local features have proven to be very successful in applications such as wide baseline matching, image retrieval, object recognition, texture recognition, robot localization, video data mining, image mosaic and recognition of object categories. Through decades'researching in this field the technology of local feature has made great progress, these approaches first detect features and then compute a set of descriptors for these features, after that, we can transform the image matching problem into feature vector measurement. However, there still exists some defects in the local feature extraction, such as high-dimensional vectors, ignoring the global information. Therefore, local feature techniques are now far from sophisticated.In this paper, a conclusion about previous achievements in local features is made. The research of this paper mainly focuses on the detection of feature points and the construction of feature description, including a short review of these techniques. Meanwhile, a novel local descriptor is presented. Additionally, we provide an instance of real application of local feature technology in image index.The main contribution of this paper are:1. After a careful analysis about the popular feature detection algorithms, we present a novel method based on salient detection, this improvement can enhance the accuracy of the interest points locality and make the local feature reflect the essence information of the image more precisely. 2. A solution that is based on colornaming, intensity value and intensity gradient is presented for the construction of the local feature. This novel local description is invariant to illumination and geometry transformation, it can also recognize the objects which are similar in texture and intensity value while different in color.3. A two-stage image index strategy based on color histogram and local descriptor is presented. This method can reduce the algorithm complexity.