基於深度學習方法在複雜光影環境下獲得掃描品質之文件影像

Abstract

現代人大多數使用手機拍攝文件影像，然而以此方式拍攝的文件影像通常會被光線影響，在文件表面產生陰影和明暗變化。這些影像可能讓閱讀者無法獲得文件的資訊，也會降低文件影像的美觀性。過去雖然有許多演算法作用在去除陰影上，這些方法其陰影去除的效果仍有所侷限，且並無前人得到具有掃描品質的文件影像。在這篇論文中，我們提出了一個使用深度學習之演算法，掃描網路，可將拍攝所得之被陰影及明暗變化影響的單張文件影像轉變成掃描品質的文件。我們的方法利用了陰影與明暗變化的不同特性設計了兩個網路，第一個網路移除陰影，第二個網路則負責校正明暗變化。我們也設計了一個產生模擬資料的方式，此方式可以快速且有效率的產生大量且多變的數據集，以訓練我們的模型。此方法所產生的資料變化量，可以訓練更為泛用的模型，以解決前人在陰影去除效果上的限制。我們的方法在陰影去除的效果、掃描文件影像品質、文件文字辨識上都達到相較於前人較好的成果。

Document images captured by smartphones and digital cameras are often vulnerable to photometric distortions, including shadows, non-uniform shading, and color shift due to the imperfect white balance of sensors. These distortions can confuse readers with indistinguishable content and background, significantly reducing the visual quality and legibility. Although real photos often contain a mixture of such distortions, most existing approaches for document illumination correction focus on a small subset of them. In this paper, we present ScannerNet, a comprehensive method that can eliminate complex photometric distortions. For exploiting the different characteristics of shadow and shading, our model consists of a shadow removal sub-network followed by a shading correction sub-network. In order to model the various illumination types in the real world, we also carefully design a data synthesis method to efficiently construct a large-scale document dataset with great variations to train our model. Extensive quantitative and qualitative experiments show that our method significantly enhances image quality by removing shadows and shading, preserving figure colors, and improving readability.