Research Of Local Invariant Feature Image Matching Algorithm And Its Application

The image matching algorithm which is based on local invariant features have gained widespread attention in recent years, becoming a hot field in image processing, 3D reconstruction, object recognition and tracking and other scientific research, and also been widely used in many fields including industrial inspection, traffic management, military fields, social activities fields and other fields. In recent years, with the rapid growth of social activities and daily needs of industrial production automation, image matching is increasingly widely used in industrial production and daily life, having got people’s more and more attention. Therefore, The research of image matching algorithm have got a fundamental significance.Image matching algorithm mainly includes pixel-based gray correlation matching, feature-based image matching, and image matching based on other theories. Image matching based on pixel gray correlation has a great image information need for processing, computing are more complex, real-time performance is poor, and are very sharp to small changes in the image, anti-noise and anti-interference ability are very poor. The feature-based image matching includes global-based invariant features image matching and local invariant features image matching. Because the global invariant feature is not readily available and the inability to distinguish between foreground and background, the image matching based on the global invariant feature is rarely used. Local image invariant features refer to local features remain the same under various transformation such as image viewpoint changes, rotation changes, scale changes, illumination changes, image blur, JPEG compression and so on. The computational complexity of image matching based on local invariant features are lower than the image matching algorithm based on gray correlation, and have strong anti-noise and anti-interference ability. Therefore, this thesis mainly studies image matching algorithm based on local invariant features. The main contents of this thesis are:Firstly, the image feature detection and the widely used methods for feature detection have Harris affine, Hessian affine, MSER, DoG, IBR, EBR, etc. We also comprehensive analyze the above several kinds of feature detection algorithm’s performance. Then describe the interest regions which have been detected, and using local invariant feature descriptor to describe these regions. SIFT descriptor and DAISY descriptor are the two most typical local invariant feature descriptor. MROGH descriptor, MRRID descriptor, LIOP descriptor and HRI-CSLTP descriptor are the most popular researches and have been paid most attention in recent years. We compared these six descriptor’s performance curves under six transformations from the image viewpoint changes to JPEG compression.Then, in order to achieve further improve of the matching effect, we improved image matching algorithm based on local invariant features. First, using the angle similarity analysis instead of the traditional Euclidean distance analysis in the matching process, while improving the angle similarity image matching algorithm, sets the minimum angle, the second smallest angle and the threshold of corresponding Euclidean distance ratio between the minimum angle and the second smallest angle. We called the improved algorithm as IBoAS algorithm. After improving algorithm, we compared the matching rate and the matching speed of the new algorithm and other known matching algorithm under six different descriptors case of six transformations from the perspective change to JPEG compression. IBoAS algorithm reduces the computational steps, effectively improve the speed of image matching, and because setting threshold to delete the mismatching points, effectively improve the match rate of the image.Finally, the IBoAS algorithm is applied to image stitching, using art building of Anhui universities, libraries of Anhui universities, south learned building of Anhui universities as the physical scene graph to image stitching applications, observing and comparing the image stitching effect.