一個新奇的兩階段熱燥點雜訊去除演算法

A Two-Stage Algorithm for Hot-Pixel Noise Reduction

吳俊霖(Jiunn-Lin Wu) ； 鄧瑋敦(Wei-Dun Teng) ； 吳豐佑(Fong-You Wu)

《科學與工程技術期刊》 4卷2期 (2008/06) Pp. 19-29

https://doi.org/10.7117/JSET.200806.0019

摘要

數位感光元件相關技術在這幾年有了突破性的發展，這使得數位相機已經廣泛地應用在各領域中，例如天文攝影等。但是相對的，其也衍生出不少問題，其中比較嚴重的是雜訊，尤其在長時間曝光攝影時，由於CCD(charge coupled device)和相機電子元件的特性所使然，更可能會以隨機斑點的形式出現電子幹擾或雜訊。曝光時間越長，此效果越明顯，此種雜訊我們通常稱爲熱燥點(hot pixels)。目前有些相機提供了雜訊抑制功能(noise reduction)來解決此一問題，但在實際經驗中發現，當曝光時間超過3~10秒時，雜訊會急速大量出現，此時雜訊抑制的效果就幫不上忙。在本研究中，我們將針對此種雜訊做詳細分析，並提出一個自動雜訊偵測演算法，並有效去除雜訊。 我們發現通常此種雜訊依其特性可以分爲兩種，一種爲近似脈衝雜訊的雜點，通常爲1 pixel大小。另一種雜訊通常散佈在紅藍頻道，大小較大，可以達數pixels。目前已有許多的線性或非線性的雜訊濾波器被提出，他們可以有效地去除雜訊，然而其卻也把影像中重要的細節也模糊化了。爲了有效去除第一種雜訊，本論文提出一混合式向量濾波器，其結合了以脈衝雜訊偵測爲基礎的向量中值濾波器與適應權重向量濾波器。此濾波器能有效復原受雜訊干擾的數位彩色影像，而達成雜訊的移除和邊緣保存。其主要的優點有：(1)所提改進之雜訊偵測演算法可以有效的降低在影像邊緣和細節部分的錯誤偵測，利用反覆的偵測雜訊而有較高的正確性；(2)所提之適應權重函數在於邊緣的像素有更爲正確的權重值並可有效的移除高斯雜訊及保存影像中的邊緣和細節部分的資訊。實驗結果顯示，本論文所提出的方法可有效的移除脈衝雜訊和高斯雜訊並保存影像邊緣和細節部分的資訊。 第二種熱燥點比一般雜訊要大得多，所以前面所提濾波器將會失敗。也有人提出利用市面上的影像處理軟體，例如Photoshop的修復筆刷工具與PhotoImapct的修容工具等，來將熱燥點拿掉，但是類似這樣的工具，都必須自己手動選取要拿掉的熱燥點並自己找材質來填補，因此耗時耗力。而本篇論文的目的則是提出一個能夠自動偵測熱燥點的演算法，同時利用影像修補的技術，將影像上令人困擾的熱燥點去除，並且維持影像中的材質與線性結構的完整。實驗結果顯示，所提演算法能有效解決數位影像的熱燥點問題。

關鍵字

熱燥點 ； 雜訊抑制 ； 向量濾波器 ； 影像修補

並列摘要

As digital cameras have become increasingly popular in recent years, taking many digital photos is now very easy. However, images with long exposure times are usually affected by a type of serious noise called hot-pixel, which comes from individual CCD/CMOS (charge coupled device/complementary metal oxide semiconductor) sensors with higher than normal rates of charge leakage. Hot pixels can appear as small pixel-size bright points of light in longer exposures, in which there might be a few bright hot pixels, several intermediate ones, and many very faint ones, forming an entire spectrum of brightness. Some professional cameras provide a noise-reduction technique used on longer exposures that involves taking a duplicate exposure with the lens covered and subtracting the image from the main exposure. However, this procedure does not work well if the exposure time is long enough.
A two-stage noise-reduction algorithm for hot pixels is proposed in this study. In the first stage, the aim is to remove both the impulse and the Gaussian noises. Many linear or nonlinear filters have been proposed to suppress or reduce noises; however, these filters also smooth out the edges and details in the image. To overcome this problem, a hybrid vector filter for color-image noise reduction is proposed in this research. By combining the impulse-noise detection approach based on both a vector median filter and weight-adaptive vector filtering for edge preservation, the proposed method achieves effective noise removal and preservation of details.
The goal of the second stage is to remove the large hot pixels in the R and B channels. The size of hot-pixel noise is usually larger than 5-to-10 pixels. In this study, an automatic hot-pixel removal algorithm based on the inpainting technique is developed. First, a noise-detection algorithm to identify the hot pixels by using a Sobel filter and morphology operations is developed. A fast exemplar-based image-inpainting approach, which achieves accurate propagation of linear structures, is then used to fill dust-spot holes in the images. Several examples of actual images are provided to demonstrate the effectiveness of the proposed method, the main advantages of which include a first-stage noise-reduction algorithm that not only removes the impulse noise and effectively suppresses the Gaussian noise but also preserves the details and edges of the image; moreover, in the second stage, the hot pixels in the R and B channels are totally removed and the texture and structure information of the image also preserved.

並列關鍵字

hot pixels ； noise reduction ； vector filter ； image inpainting