| Abstract [eng] |
In this research we propose a solution to create 3D model using unordered scanning electron microscope (SEM) images. These images are very popular in medicine, chemistry to display specimen’s surface. In order to reconstruct accurate 3D point cloud, you need to execute two steps: structure from motion and multi view stereo. In structure from motion process solves image position in the 3D space problem to get extrinsic (image translation, rotation), intrinsic (camera) parameters and reconstruct sparse 3D point cloud. In multi view stereo process using extrinsic and intrinsic parameters we generate depth maps using input images. Next these depth maps are fused to create a dense 3D point cloud. In later steps it uses algorithms to remove some noisy cloud points using filters and then we generate surface map to fuse with point cloud points. The main problem in this research is to get extrinsic camera parameters. The accuracy of these parameters depends on image quality and extrinsic parameters algorithms. One of the biggest problems to reconstruct accurate 3D point cloud are input images. It is very difficult task to acquire a perfect SEM image because it needs to employ electrons instead of light that uses phones or cameras to determine specimen’s surface. SEM images usually have noise, darkness, brightness, illumination, haze, blur and so on. So, we analyzed methods to enhance SEM image quality. Extrinsic camera parameters accuracy depends on quality and quantity of feature points (inlier), results of feature matches between two images and its calibration. Other feature points - outliers which come from noise and false matching and have big impact on accuracy of 3D SEM reconstruction. We analyzed existing feature detection, feature matching algorithms and matches removal algorithms and other things. Also, we use previously modified images and analyze how feature matching accuracy improved. The experiments are presented at the end of the thesis and advantages and disadvantages are evaluated based on the 3D model quality. |
