An Application of Machine Learning for De-interlacing Technique

• Main Authors: Han-Hui Hsiao, 蕭涵徽
• Other Authors: J. H. Jeng
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/03544343601638200869

碩士 === 義守大學 === 資訊工程學系碩士班 === 94 === 　　Interlaced sequence is the major way for storage and data transmission in many video formats and TV systems. Each frame in the interlaced sequence is composed of two field signals, and there exists a time interval between them. Due to this, sawtooth effect may be generated during the process of transforming interlaced sequence to progressive one. Therefore, a de-interlacing method not only transforms the signal sequence, but also eliminates the undesirable visual effects with other algorithms for better video quality. 　　Traditionally, de-interlacing is applied to a given field signal, and a reconstructed frame is obtained from the evaluation of linear interpolation and multiple edge direction prediction. The mostly used edge direction prediction method is Edge-based Line Average (ELA), and the proposed method with noise-tolerant in this paper is improved on the basis of ELA. Additionally, the de-interlacing method based on machine learning is another emphasis of this paper. The proposed method is entirely different to traditional linear interpolation. It possesses accurate prediction capability through excellent learning mechanism of Support Vector Machine (SVM). In the experiment, different kernel functions are used to investigate the de-interlacing results. In order to obtain better de-interlaced image quality, training data and testing data are preprocessed in advance: Edge Direction Invariance is used to improve the sawtooth effect, and Brightness Invariance is used to suppress the raise of estimate values. The experimental results show that a de-interlaced image with smooth edges is obtained, and both the image quality and the visual effect are in reasonable results.