基於不同圖像條件的去模糊方法

Abstract

圖像在各領域中被廣泛的使用，大家對於圖像品質的要求也越來越高，但在拍攝照片的過程中，我們難以避免手部的晃動或是目標物體的移動，而導致圖像產生模糊。想要使模糊圖像變得清晰我們就必須估測出模糊內核，但在一般情況下，模糊內核與清晰圖像都是未知的，使得影像去模糊會是一個ill-posed的問題，為了解決此問題許多研究方法不斷被提出。近年來，大多數的去模糊方法都是基於MAP架構，他們從自然圖像中學習有用的圖像先驗知識來去除圖像模糊，並且假設內核具有稀疏性，這些方法使得圖像去模糊取得了重大進展。在本文中我們主要針對那些假設圖像模糊是均勻且空間不變的方法來進行討論，我們比較多種基於不同圖像先驗的去模糊方法，包括補丁先驗，梯度先驗和像素強度先驗等等，並且從各演算法中推測可能導致內核與潛在圖像估測不精確的原因，此外我們利用各種不同場景的模糊圖像來比較各方法的性能。由於圖像先驗的設計是基於自然圖像中的統計特性，因此那些與統計特性不相符的圖像的去模糊結果會受到很大的影響，另外透過不同範數來約束內核的稀疏性也會影響去模糊的結果。 在實驗結果中，我們利用合成與真實圖像來測試各方法的性能並且調整不同的參數來觀察對結果的影響，我們發現有些方法對於內核的大小與參數的設定非常敏感。在定量評估上，我們利用平均結構相似性指數（MSSIM）和峰值信噪比（PSNR）來對恢復的合成圖像和內核進行評估。

As images are widely used in various fields, our expectation toward the quality of images are getting higher accordingly. However, it is difficult to avoid hand shake or movement of the target object during the photographic process, resulting in blurred images. In order to make the blurred image clear, we must estimate the blur kernel. Nevertheless, the blur kernel and clear image are unknown under normal circumstances, making the removal of the blurriness a matter of the ill-posed issue. To solve this problem, many methods have been developed. In recent years, most of the methods are based on the MAP framework. They learning useful image prior knowledge from natural images to remove image blur, and assume that the kernel is sparse. These methods have made huge progress in image deblurring. In this thesis, we are discussing the issue under an assumption that image blur is uniform and spatially-invariant. We compare several deblurring methods based on different image priors, including patch priors, gradient priors, pixel intensity priors, etc. Then we speculate on the cause of the inaccuracy of the kernel and the latent image. In addition, we compare the performance of these methods by blurred images of different scenes. Since the designs of these image priors are based on statistics of natural images, the deblurring results of those images that do not match the statistical characteristics will be greatly affected. Moreover, using different norms to constrain kernel sparsity will also affect the final result. In the experiment, we used synthetic and real images to evaluate the performance of each method and adjust different parameters to observe the impact on the results. We found that some of the methods have high sensitivity toward the size of the kernel and parameter setting. In quantitative assessment, we used Mean Structural Similarity (MSSIM) and Peak Signal-to-Noise Ratio (PSNR) to evaluate the recovered synthetic image and kernel.