Study On The Non-uniform Illumination Image Registration Based On Fast Feature Extraction

Image stitching technology stitches and fuses multiple images with overlapped regions which taken from different directions of the same scene to produce a panoramic image with large field of view,high resolution and natural sense,which is mainly applicable to fields such as medical imaging,satellite remote sensing,traffic monitoring,military security and other scenarios that require capturing large areas.With the expansion of panoramic image application requirements,image stitching technology also faces some challenges.The main difficulties can be attributed to issues such as unclear image details and fuzzy features caused by non-uniform illumination,as well as the time-consuming problem of feature extraction for high-resolution image of complex target scene.In order to achieve more accurate and natural registration of non-uniform illumination images,a fast adaptive local enhancement algorithm combines linear and nonlinear correction was improved in this thesis.At first the input image was adaptively linearly enhanced in saturation,and nonlinearly corrected in brightness based on improved Retinex algorithm and optimized 2D-Gamma function.Then the above algorithm was optimized based on CPU thread parallelism to realize the fast adaptive local enhancement of brightness and saturation in nonuniform illumination images.Meanwhile,the enhanced images were evaluated subjectively,and the images’ quality was measured objectively by metrics such as information entropy(H),standard deviation(SD),average gradient(AG)and blind/referenceless image spatial quality evaluator(BRISQUE).The experimental results show that the proposed algorithm has obvious advantages in enhancing image detail texture,increasing the number of feature points,preserving image color fidelity and balancing brightness and saturation.To address the slow speed issue of traditional feature extraction algorithms,a speeded-uprobust features(SURF)algorithm based on CUDA acceleration optimization was improved in this thesis,with the aim of meeting the real-time requirements of high-resolution image stitching.The feature points detection and description steps of traditional SURF algorithm were accelerated in parallel based on CUDA from the aspects of memory access,program design,and algorithm optimization.The matching accuracy was improved by combining coarse and fine matching and adopting bidirectional matching strategy.The experimental results show that the proposed parallel algorithm can achieve an acceleration ratio of more than 10 times compared to serial algorithms for images with different resolutions.Moreover,bidirectional matching accuracy was improved by 17% compared to traditional FLANN unidirectional coarse matching + RANSAC fine matching algorithm,with the best accuracy reaching 96%.In order to effectively reduces the problem of dislocation and ghosting in multi-view images stiching,an image stitching process based on grid deformation was optimized in this thesis.In this thesis,CUDA-based feature extraction algorithm was performed on pre-processed image after non-uniform illumination correction.Then the optimal matching points were obtained based on sequential RANSAC algorithm.Next the grid model with adaptive division and topology structure were constructed,and the protection item of linear structure was added to constrain the grid deformation,so as to complete the grid-based image registration.Finally,the panoramic image was obtained by mapping transformation and linear weighted blending.The experimental results show that the proposed algorithm is more suitable for the natural stitching of non-uniform illumination images compared to grid stitching algorithms such as APAP,AANAP and NISw GSP.