Fisheye Image Correction And Its Comparisons

Compared with a flat lens,a fisheye lens has a shorter focal length and a larger viewing angle,and it has a wide range of applications in photogrammetry,medical applications,national defense and military,security surveillance and other fields.Due to the severe distortion caused by the distortion,the fisheye camera image needs to be corrected to a perspective projection image.This thesis focuses on the problem of distortion correction of fisheye camera images.The main work is as follows:(1)Analyzed and summarized the research on fisheye image distortion correction,compared the difference between the pinhole imaging model and the fisheye imaging model,distinguished the flat lens image correction method and the fisheye image correction method,and discussed the correction process the various coordinate systems involved,correction projection methods and distortion models,etc.(2)Select the scanning approximation algorithm,improved scanning approximation algorithm and area statistics method to extract the effective area of fisheye distortion image and compare the experiments.Choose a better algorithm as the effective area extraction algorithm of fisheye image in this thesis.On this basis,two traditional model-based fisheye image correction methods are realized,latitude and longitude mapping and equidistant projection.The image correction results of the two methods are given and the two traditional methods are analyzed and compared.(3)Two deep learning fisheye image correction methods based on generation and regression are realized.The two methods are trained through the open source simulation data set,and the image correction results of the two methods are given.(4)The experimental data is obtained by setting up three experimental environments of different image scene complexity,different image effective area shapes and different image distortion degrees.With the aid of two indicators of peak signal-tonoise ratio and structural similarity,the above four methods are further compared,and the following are obtained Conclusion: 1)When the complexity of the image scene is relatively light and intermediate,the fisheye image correction method based on generated deep learning is better than the method based on regression,and when the complexity of the scene is high,the method based on regression is the best;2)When the effective area of the image is a circle,the traditional distortion image correction method is better than the method based on deep learning,and the latitude and longitude mapping method is the best.When the effective area of the image is square,the fisheye image correction method based on deep learning is better than the traditional method;3)When the degree of image distortion is different,the longitude and latitude mapping method and the deep learning method based on regression have the highest stability;4)Fisheye lens The size of the field of view will not affect the correction accuracy of the image distortion algorithm.