Despite great public interest in cultural assets, knowledge of these assets is often restricted because of deterioration and collapse. Consequently, it has become an important goal in the computer science community to model and record such assets. This digital data is then used to create digital media such as, computer graphics (CG), and virtual reality (VR) content. Furthermore it is also used for restoration and preservation analysis.

For modeling cultural assets, it is important to understand how to preserve the important information about them, including shape and surface color. Usually, shape information can be obtained by a laser range sensor, and this information has become more accessible with improved data processing algorithms. On the other hand, current color imaging systems like digital still camera (DSC) are usually represented with the traditional red, green, blue (RGB) color model. RGB cannot always provide accurate color information. The color of images captured with DSC is dependent on both the characteristics of the device and the condition of the illumination environment, making it difficult to accurately represent color appearance in the real world. Consequently, we need to capture surface spectral reflectance as ultimate color information in order to preserve the accurate color of cultural assets.

Spectral reflectance is inherent in the nature of objects. Object analysis using the fact that different materials have different spectral reflectance is performed in many fields, such as medical imaging, agriculture, archaeology, and art. However, in the real world, it is difficult to obtain and handle a multispectral image effectively. Cultural assets are often found in an outdoor environment or in a dark environment can pose different problems. In an outdoor environment for instance the temporal alteration of the illumination environment changes greatly. This causes effects of saturation and underexposure when measuring the spectra. Moreover, most cultural assets have at their surface complex reflection, absorption, and scattering, with a color mixture between the top and bottom layers, making material segmentation impossible. This paper targets analysis of cultural assets having such multilayered characteristics.

Our goal is development of novel multispectral imaging systems for modeling cultural assets, proposing reflectance analysis methods using a multispectral image, and applying them in practice. The paper proposes the following three tasks related to preservation, release, and analysis.

First, in order to make VR contents by using 3D data, texture images by DSC are generally used as color information. However, the color of images captured with DSC is dependent on both the characteristics of the device and the condition of the illumination environment. A fundamental problem here is that the color information is not accurate. In a narrow environment, such as a tumulus, compact mobility is necessary to measure spectra. For these circumstances, we propose a color restoration method that uses both high resolution images captured by DSC and spectral information captured by a conventional spectrometer. This is a practical method from the viewpoint of automation and computational cost.

Second, we propose an efficient acquisition and segmentation method of multispectral images in outdoor environments. A conventional multispectral imaging system may have two kinds of cameras. The first is a multiband camera, which is mainly used in the color reproduction field and does not have high spectral resolution, but has high image quality. The second is a hyperspectral sensor, which is mainly used in the aerial remote sensing field and does not have high image quality, but has high spectral resolution. Compared to these systems our solution has not only high image quality and sufficient spectral resolution for object analysis but also a wide capture angle. The multispectral image segmentation method can handle object surfaces having complex reflection based on a statistical procedure.

Finally, many object surfaces such as wall paintings are composed of layers of different physical substances, and are called "layered surfaces." Such surfaces have more complex optical properties than a diffuse surface and are generally unable to be segmented. Our method proposes a novel physical model we call the Spider model, which decomposes optical properties.

The main contribution of this thesis is that the author proposes two multispectral imaging methods and two reflectance analysis methods which are applied not only in theory but also in practice to show their viability. The contribution can be specifically summarized by the five following points: First, we have developed a practical color restoration method based on spectral information for making VR contents, have actually produced VR contents by using restored images, and have also released them in Kyushu national museum. Second, we have developed a multispectral imaging system that can efficiently acquire spectra in a wide field. Third, we have proposed a multispectral image segmentation method based on statistical procedures. Fourth, we have proposed the Spider model as a physical model for layered surfaces, and have also proposed decomposing complex reflection components of a layered surface. To our knowledge, there are no methods sharing our goals and techniques. Finally, we have applied our methods to both the reflectance analysis of tumuli and the spectral analysis of bas-relief in the Inner Gallery of the Bayon Temple. These methods are specifically designed for modeling cultural assets, but they can be used in other fields as well.

本論文は,「REFLECTANCE ANALYSIS OF LAYERED SURFACES USING A MULTISPECTRAL IMAGE(マルチスペクトル画像を用いた層状表面物体における反射率解析)」と題し,マルチスペクトル画像を用いて層状表面物体の反射率を解析する研究をまとめたものであり,全五章で構成され,英文で記述されている.

第一章は,「Introduction」と題し,スペクトル情報の取得と反射率解析に関する研究例を紹介している.また,研究目的,論文の構成についても述べている.

第二章は,「Color Restoration Method Based on Spectral Information Using Normalized Cuts」と題し,デジタルカメラ(DSC)とスペクトロメータを組み合わせることで、簡便で正確に多くのマルチスペクトル画像を復元する手法を提案している.三次元データを用いたVRコンテンツ制作において,色情報は主にDSCにより取得されるテクスチャ画像が用いられるが,DSCより得られるRGB画像は撮影環境や機器の特性に左右されるため,正確な色情報を公開することはできない.また、古墳のような狭い環境下では,撮影において機動性が求められることをから、機動性に優れたスペクトロメータのスペクトルをRGB画像に適用することで問題を解決している.さらに手法により復元したマルチスペクトル画像を用いてコンテンツを制作することで手法の実用性を示している.

第三章は,「Multispectral Imaging for Material Analysis in an Outdoor Environment Using Normalized Cuts」と題し,屋外環境において広範囲のマルチスペクトル画像を効率よく取得するシステムの開発と,物体を解析するためのスペクトル画像の領域分割手法を提案している.従来のマルチスペクトル画像取得システムでは主に以下の二種類が使われている.一つは、画質は良いがスペクトル解像度が低い色再現の分野で利用されているもの,もう一つは,スペクトル解像度は高いが画質が悪いリモートセンシングの分野で利用されるものである.提案システムでは画質の良さと物体解析に十分なスペクトル解像度の高さの両方を備え,さらに広い範囲の画像を効率よく取得できるシステムを開発し,屋外環境において急速に変化する照明環境に対応した適応的露出推定法を提案している.また,マルチスペクトル画像の領域分割手法では層状表面の複雑な反射特性を加味した統計学的手法が提案され、他の手法との比較により手法の有効性を示している.さらに,バイヨン寺院の着生物解析に手法を適応することで手法の実用性を示している.

第四章は,「Decomposing Complex Reflection Components of a Layered Surface Using the Spider Model」と題し,層状表面の光学特性を表現する物理モデルとしてスパイダーモデルを提案し,これを用いて層状表面の領域分割及び各層の光学特性を推定する手法を提案している.壁画などの物体表面は下地と表面層で構成される層状表面である.このような表面は両層が混ざり合った複雑な反射特性を持つため,表面層の領域のみを従来の領域分割手法などを用いて取得することは難しい.そこで手法ではこのような層状表面の光学特性を表す新しい物理モデルとしてスパイダーモデルを提案し,このモデルを用いて表面層領域を正確に領域分割するだけでなく,表面,下地層のスペクトル情報と上層の吸収特性に分解する手法を提案している.

第五章は「Conclusion」と題し,本論文の成果を要約すると共に今後の課題が示されている.

以上これを要するに,本論文では,マルチスペクトル画像の取得のための,高効率なシステムと高精度なシステムを開発し,これから得られるマルチスペクトル画像を解析する2つの手法を提案し,これらの手法を実際の文化財に適用し,文化財の表面の着生生物解析や顔料解析への道を提示し,また得られたスペクトル画像を用いてVRコンテンツの作成例を示すことで,これらの技術が、文化遺産のディジタル保存に有効であることを示したもので,電子情報学上貢献するところが少なくない.

よって本論文は博士(情報理工学)の学位請求論文として合格と認められる.