Research On Raw Data Matching Based On The Binary Description

Image matching plays an important role in the applications of computer vision as a key technique. It is the basis of the advanced applications, such as object tracking, intrusion detection, visual search, and 3D reconstruction, which is widely used in the fields of security monitor, intelligent transport, medical diagnosis, and remote sensing. It has drawn much attention from past to present. Recently, the image matching algorithms based on the binary description are becoming the focus for the reason that they utilize simple and effective feature description leading to much reduced time cost compared to the methods using floating descriptors.However, under the tendency of high-resolution imaging, the requirement of real-time matching in the practical applications is increasingly enhanced. These approaches based on the sequential framework which completes the analysis after the imaging stage are difficult to break the bottleneck of efficiency. Therefore, in this paper, the raw data is brought in for a whole new parallel processing framework, and the algorithm parallel and integrated matching for raw data(PIMR) is introduced aiming at radically solving the efficiency issue. Experiments demonstrate that comparing against the state-of-the-art approaches, the time cost of the proposed method PIMR is reduced dramatically, and meanwhile it could gain comparable precision in most cases which is robust to general image transformations as well. The work in this paper is not only the basis of fast image matching, but also the preliminary study on the raw data. Generally speaking, the research is of theoretical and practical significance.The main contribution of the research is explained as the following four aspects.(1) The local feature based image matching methods have been surveyed mainly from the aspects of feature detection, description and matching, of which the ones including BRIEF, ORB, BRISK and FREAK have been studied closely. As the sequential framework employed in these methods restricts the improvement in the efficiency, a parallel processing framework is hence designed to make the two stages which are the imaging and the processing parallel using a multi-core processor.(2) The realization of the parallel processing framework benefits from the raw data. We have learned the formation, property, data format of raw data, and find that it is more appropriate for analysis than digital image since it maximally keeps the information of the real scene. Therefore, raw data is introduced to be the object of the parallel framework, and a fast matching algorithm coined a parallel and integrated matching for raw data(PIMR) is presented in the paper. One thread of the approach is responsible for the demosaicing, and the other one is for realizing image matching by raw data analysis.(3) Raw data cannot be directly used for further processing because of the lack of pixel information, thus raw data reconstruction operation is utilized to fuse the color information effectively. The operation which is easy and fast adopts the pixel information within a 2×2 unit to reconstruct the intensity of the target pixel to make raw data more proper for the matching.(4) The validity of PIMR is tested with respect to the precision and time cost. Firstly, acquire the raw data test sequences from the Affine Covariant Features dataset. Secondly, compare the proposed algorithm against the state-of-the-art methods on the raw data dataset, including the recognition, recall-(1-precision) curve and time cost. Experiments show that the proposed PIMR is robust to image rotation, illuminance change, image blur and compression whose recognition can reach above 99.5%. Moreover, compared to the ORB, BRIEF, BRISK and FREAK, the time cost of the proposed method is reduced significantly. Especially, the average time of PIMR is nearly 5 times faster than BRISK for the wall sequence.