Research On Image Stitching Algorithm Based On Saliency Seam Fusion And ASGD Keypoint Extraction

Image stitching is an important research direction of computer vision.After years of development,this technology is widely used in the fields of satellite remote sensing,medical imaging and quality inspection.The keypoint extraction stage extracts image feature information,and then generates specific descriptors that can represent the keypoints.The image alignment stage matches and filters the descriptors to obtain single or multiple matrices,and aligns multiple images with overlapping areas to the same plane through transformation.The image fusion stage combines the transformed images to be stitched into one image and eliminates the stitching traces in the overlapping regions to make the stitched result better subjective.The following aspects are studied for the image stitching characteristics and problems:(1)we proposes a feature extraction algorithm based on adaptive thresholding quadtree with fused gradient differences(ASGD).In the feature extraction stage,the ASGD algorithm divides the image into multiple subregions,and extracts feature points in the subregions separately using the FAST.Because the subregions contain different information among them,different FAST thresholds should be set according to the actual conditions of the subregions.The ASGD algorithm determines the FAST threshold for a subregion by calculating the average pixel fluctuation of that subregion.The number of feature points extracted by the above methods is often larger than the target,so the ASGD algorithm optimally filters the point set with the help of the quadtree method.In the descriptor generation stage,the ASGD algorithm adds gradient comparison to the original grayscale comparison to enrich the information contained in the descriptors.To make the descriptors more robust,the ASGD algorithm obtains comparison pairs with different dimensions by constructing concentric circles in the neighborhood of feature points.Too high descriptor dimensionality can seriously affect the efficiency of the algorithm.The ASGD algorithm defines distance thresholds to filter the short distance subsets used for comparison,and also uses circular comparison of grayscale and gradient to reduce dimensionality.(2)For the parallax problem caused by the difference of shooting angle and depth of field,we proposes the optimal seam fusion algorithm based on visual saliency.The human eye has a limited ability to distinguish colors below a threshold.For this reason,we uses an excitation function to process the energy spectrum and enhance the distinction.Also because the human eye always focuses on the salient regions in an image,distortions and artifacts in the salient regions are more noticeable.The algorithm adds saliency features to the seam search strategy,specifically.First it calculates the saliency weights of the overlapping regions using the MBS salient target detection algorithm.Then it weights the energy spectrum as a way to enhance the energy of the salient regions,and finally searches for the best seam using the graph cut algorithm.The experiment show that the stitching seam can effectively avoid the salient regions in the image,and the final stitched result is more natural to the human eye perception.(3)The algorithm is implemented on windows platform using computer vision library open CV and C++ programming language.Experiments are made on public datasets such as SEAGULL,Parallax,photomicrograph dataset of rocks for petrology teaching,blood cell dataset and satellite images dataset of water bodies.The robustness and effectiveness are verified by comparing the degree of uniform distribution of keypoints,matching accuracy,cost of unit descriptor,image transformation quality and seam quality under different scenes.