새로운 물체 검출 기법을 이용한 효과적인 노이즈 제거 알고리즘

Abstract

Digital images, which are a subset to digital signals, are normally contaminated by many types of noise during the steps of image capturing. Noisy digital images and videos degrade human and machine perception for image processing systems. Therefore, Image de-noising plays an important role in image processing, computer vision, and pattern recognition fields. The main goal of image de-noising is to enhance or restore a noisy image and help the other system or human to understand input images better. In this thesis, we propose efficient approaches for image and video de-noising using novel object detection methods. The reason why various image de-noising algorithms use object detection based methods, is that the human visual system (HVS) tends to pay more attention to bright and high-focused region and less attention to dark and low-focused region. Thus region adaptive de-noising algorithms improve the performance of reducing noise in terms of both subjective and objective image quality assessment. In the first proposed method for a still image, the efficient region-based noise removal algorithm which segments an input image into two region, object and background, and applies adaptive de-noising algorithms suitable for each regions, are proposed. Most previous methods depend on the assumption that the images are in noise-free conditions, which leads to high false positive rates in noisy images. However, we introduce a novel image segmentation algorithm called, adaptive second-order-statistics (ASOS) for noisy low depth-of-field (DOF) images. We derive the following three solutions for reliable ASOS-based image segmentation which indicates the spatial distribution of objects in the face of noisy images: • The spatial distance weight: Within neighboring pixels, distant pixels have lower co-relations than close pixels. Therefore the spatial distance weight gathers adjacent pixels into groups for increasing accuracy near edges. • The tonal distance weight: Within neighboring pixels, the same regions tend to have similar tones. Therefore the tonal distance weight gathers similar pixels into groups for clarifying the edges of object and noises. • The second-order-derivatives (SOD) weight: This weight, representing the ratio between the SOD value of the pixel and the maximum value across the whole image, let striking artifacts considered as noise pixels or out-focused pixels removed in advance. In the second proposed method for a video, two method to enhance the accuracy of noise removal are proposed, one is the matched point-related method and the other is the adaptive spatial-temporal reference filter. The matched points which trace same pixels within consequence pixels, makes reliable reference pixels in the neighboring frames. Unlike previous algorithms, the proposed method adaptively selects reference pixels within spatially and temporally neighboring frames. These selections are based on the voting scheme among the spatial reference, the temporal reference, movement-awareness temporal reference and the others. Voting criterion is the summation of weights within each reference pixels. Thus the proposed method uses reliable information about matched pixels on neighboring frames. As a result, the proposal enhances the accuracy of video de-noising. By virtue of these algorithms, diverse requirements of imaging systems on image de-noising quality can be satisfied. Segmentation accuracy in noisy image is improved by 25.4% comparing to conventional method. De-noising accuracy of the still image is improved by 11.34% comparing to fast bilateral filter maintaining computational times. De-noising accuracy of the video is improved by 6.25% comparing to the conventional. In conclusion, the proposed methods produce cleaner and sharper images than conventional algorithms.