Research On Key Technologies Of Parallelized Aerial Image Fusion

Aerial photography becomes popular with the popularity of drones such as DJI.Aerial photos have larger resolution and field of view than traditional photos.This brings challenges to image processing techniques for aerial images.Aerial image fusion is an important technique and plays a key role in high quality aerial image generation.While traditional image fusion approaches cannot handle aerial images,we study the par-allelization of state of the art fusion approaches,by making them fully parallell and implementing them in GPU.As a result,the performance improves two orders of magni-tude.In summary,we have the following contributions:1.We proposed a stitching model for aerial images.Traditional image stitching mod-els assume that the input images are captured at the same viewpoint,while aerial images are in fact captured in different points of time by a flying aircraft.We pro-posed a multi-view based model to handle aerial images,and experimental results validated the effectiveness of our method.2.We improved the traditional image fusion algorithms.To deal with the low effi-ciencyproblemofmulti-bandfusion approach and mean value coordinate(MVC)fusion approach,we proposed to use GPU-acceleration technique to improve their efficiency.Our approach improves the original multi-band fusion algo-rithm by using equivalence-weight functions.In this way,the laplacian pyramid decomposition can be computed in parallel.When implemented in GPU,our pro-posed approach is 200 times faster than the original approach.We proposed GPU-accelerated mean value coordinate(MVC)fusion approach.MVC fusion is an approximation of Possion fusion,and it avoids solving large linear functions.In MVC fusion boundary differences should be calculated first,and this brings challenge to parallelization,especially when there are multiple regions to be blend-ed.Our approach first calculates the boundary differences in an efficient way,then the value of each pixel to be calculated can be computed in parallel.When imple-mented in GPU our approach is 80 times faster than the original MVC fusion.3.We proposed an improved fusion approach that can avoid bleeding artefacts.This approach is based on MVC fusion and the improvements are in two folds.Firstly we improve the way of calculating the boundary differences,to make it more robust.Secondly instead of fusion target region to fit the source region we adopt two-way fusion that both source region and target region change to fit the other.Compared with the original MVC fusion,Our approach improves the visual quality of the fusion results.