學習一卷積神經網路於去馬賽克研究

Abstract

本論文提出一個基於卷積神經網路的去馬賽克演算法，方法主要在學習馬賽克影像和原始影像之間的關係。和目前表現最好的一些演算法比較後也證實了這種由大量資料自動學習出來的特徵更能有效率處理去馬賽克問題，而且實驗結果幾乎勝過目前的演算法。我們也將此架構應用於更具挑戰性的議題上：處理同時具有顏色和曝光度感光元件的去馬賽克問題。此種感光元件能夠達成單張影像生成高動態範圍的效果，實驗結果也顯示我們的方法有優異的去馬賽克效果。 此外，我們發現大部分文獻都只針對濾色片設計或去馬賽克演算法進行研究。此論文因此提出了一個新的神經網路層—濾色片設計層，加入原本去馬賽克卷積神經網路而形成一新卷積神經網路。該卷積神經網路能夠同時學習濾色片和去馬賽克方法，實驗結果驗證了此卷積神經網路的去馬克效果效果優異，在峰值信噪比和視覺呈現上甚至超越了我們先前結果。

This thesis presents a demosaicking method based on the convolutional neural network (CNN). Our method learns an end-to-end mapping between the mosaic samples and the original image with full information. The comprehensive evaluation with 10 competitive methods on the popular benchmark confirms that the data-driven, automatically learned features from CNN can be more effective and the proposed method outperforms the current state-ofthe-art algorithms. The proposed framework has also been proved effective for a more general but challenging task, spatially varying exposure and color (SVEC) demosaicking, for reconstructing an HDR image from a single shot. In addition, based on the observation that previous literatures usually develop their algorithms on pattern design or demosaicking separately. In this thesis we combine these two problems and form an optimization problem with CNN by introducing a new layer, pattern layer. Experiments show the better performances than Bayer-CFA-based demosaicking CNN on both PSNR value and visual quality.