| Abstract [eng] |
In this research, in the context of image analysis and processing, contour recognition methods are investigated for solving the problem of road surface defect recognition and lateral chromatic aberration in eye fundus images. The recognition of object contours in raster images is relevant in many practical fields: astronomy, meteorology, forensics, warfare, industry, engineering, medicine, archeology, robotics, and many others. Different methods for recognizing contours are based on differential operators, energy functions, hysteresis, neural network methods, etc. Such methods are effective in images with low color variations, with a lot of uniform textures or objects that are always symmetrical in relation to each other. But in reality, such images are rarely seen, and the accuracy of the methods developed is greatly reduced. In this thesis, images such as road surface defects or timber knots which do not have a uniform texture, have large color variations, and objects that are always in different shapes, are analyzed. The thesis presents a method for identifying contours of road surface defects and timber knots, which, by applying the binary threshold method, evaluates each point of the contour or noise by weight and uses the Dijkstra algorithm to find the most likely contour. In order to further reduce the points that are not recognized correctly, the most distant points are spotted and removed. In this study, this method was compared to other analyzed contour detection methods. Following the contour detection experiments with road surface defects, it was observed that the active contour technique only recognizes contours with an accuracy of 53% when compared to the accuracy of contours detected by the human eye. The new contour detection method introduced in the thesis is 25% more accurate than the active contour method when recognizing the contours, when both are compared to the human-defined contours. The precision is measured by the difference in the total human-defined contour area with the human-defined and method-calculated contour areas. Another topic investigated in the research for which contour detection methods were used was the reduction of lateral chromatic aberration. Chromatic aberrations occur when white light passes through the optical lens system and dispersion results in various distortions of the image. As the wavelength of each color is different, the angle of deviation of the light radius differs from one color to another after passing through the lens. All cameras use optical lenses, and various distortions occur in all optical systems. In order to improve the quality of the image captured, it is essential to reduce chromatic aberration. The development of mechanical methods based on the adding of achromatic, apochromatic or diffractive lenses to the optical system helps to reduce the effect of chromatic aberration. However, when limiting the size of the camera, mechanical methods of solving the problem are not suitable due to the occupancy of space and other reasons. This problem can be solved using software methods. However, due to a variety of optical system layouts, it is not possible to create one universal method for eliminating chromatic aberration, so this problem needs to be addressed for each optical system individually. An analytical review has shown that the lateral chromatic aberration effect can be reduced by methods based on calibration patterns, but such a solution is only appropriate when the camera, with which the scene was captured, is available. To reduce the effect of lateral chromatic aberration without having the camera with which the scene was captured, a new method for reducing lateral chromatic aberration in eye fundus images is presented in the thesis. It is based on the detection of contours in the blood vessel network and does not require the camera with which the image was captured. |
