Research And Implementation Of Fast Image Mosaic

There are a variety of methods to composite a series of pictures to a large image. Fast image stitching is one of the best choices whose major role is to quickly splice a certain number of small pieces of the images into a large area of the panoramic image by software technology. Fast image mosaic is mainly used for real-time and sudden emergent situations which require the algorithm is higher utility, such as knowledge of the disaster after it happened, view of the substantially ground landscape by aerial pictures, warning of the natural disaster, etc. These applications show that making research on fast image stitching technology can not only improve the mosaic theory, but also solve practical problems.The objective of fast image stitching is to acquire mosaic image quickly and accurately, the difficulty of the technology is how to obtain accurate registration and reduce image stitching error. Based on study of the existing fast stitching algorithm which is low real-time, poor quality stitching, less practical and other issues, we make research in the characteristic binary descriptors, fast and accurate registration, stitching practicality and other related technologies. The main contents and contributions are as follows:1. Speed-up robust features(SURF) points are complexity and it's time-consuming to compute the main directions and feature descriptor. It also needs larger space to store feature descriptor than binary features whose computation is efficient. In this thesis, we propose a binary feature detecting algorithm BSURF(Binary SURF). First, we extract features on the multi-scale image, and then obtain the characteristic direction by using the gray information. The final, we use binary feature vectors to instead the floating-point descriptors. We use many quite different images under several scenes to test the proposed algorithm. The experiments demonstrate that the algorithm can get better performance characteristics, use less time consuming and have greater efficiency than SURF algorithm.2. The error of stitching mainly comes from the registration stage. Aiming at the problem that the traditional registration method is time-consuming and the registration result is not accurate, we improve the registration model by selecting the best match points. Our algorithm uses a “two-way ratio method” and an improved RANSAC algorithm to select double. After calculating the optimal registration model, we use a nonlinear optimization algorithm to optimize the model. We use various audiences under different circumstances to test the property of our arithmetic. It is proved that the proposed algorithm reduces the registration times significantly, and our matching points are more reasonable and accurate.3. If we fuse the image directly after splicing, it would leave visible traces of stitching which absolutely results in poor fusion. Therefore, we use a hybrid approach to improve the fusion effect. We combine the Laplacian pyramid multi-band fusion algorithm with the brightness of balance and play their advantages together. We use various objects under different light sources to do a large number of comparison experiments. The results indicate that the algorithm has obvious advantages and the resulting image is seamless and color-continuous. The algorithm achieves a natural fusion.4. Traditional stitching implementation has a defect of the less practical, we present a stitching method based on seam called “dynamic line” type of mosaic which can realize to output the tile diagrams and input images at the same time. The new algorithm improves the practicability. Through the same conditions experimental comparisons, our method has greater efficiency than traditional methods and the service conditions are simpler. The results demonstrate that our algorithm is more practicality.In summary, this thesis makes research on fast image mosaic deeply, proposing a BSURF feature extraction algorithm to detect image features. Using an improved registration algorithm in the feature match to reduce the time of obtaining registration model and the model errors. In the image fusion, combining the brightness balancing strategies with multiband fusion to depth fusion. Finally, we design a real-time stitching algorithm implementation. A large number of comparative experiments conclude that under the same objective conditions, our method has a higher efficiency, better quality and greater stitching practicality.