Similar to the human eye, RGB cameras have three type of sensors, each sensitive to a different and rather wide range of visible light wavelengths from about 380nm to 780nm. Therefore, the response of the sensor is comparable to the human color sensation, but lacks the details of the scene reflectance values at each wavelength due to the summation process. In some applications, such details happen to be crucial for extracting certain information.
Instead, hyperspectral imaging provides up to hundreds of narrow-band sensors that measure the light in steps of only a few nano-meters. This extra information is useful in many applications such as detecting the materials, checking the production quality, and providing contrasts that are hidden in a typical RGB image. These applications make hyperspectral imaging a valuable tool in remote sensing, cultural heritage preservation, industrial quality control, and diagnostic medical imaging among others.
However, the drawback is the cost of this imaging method in terms of the device availability and processing time. A middle ground method is to use multispectral cameras, with a few (up to tens) of sensitivity bands, and estimate the hyperspectral data using a priori data and statistical models.
This thesis contributes to the pipeline of hyperspectral imaging, from the image acquisition procedure to the estimation model and the processing algorithms. It discusses the best practices to acquire images with acceptable signal-to-noise ratio, particularly in complex cases where the dynamic range of the scene is high. Then, it proposes and tests a number of link functions to improve the estimation quality of spectral images. Finally, it proposes a spectral feature extraction method based on mathematical moments. Features are often numerical properties that characterize the phenomenon under study and are used in automatic pattern recognition algorithms, a process that is time-consuming or difficult if done manually.
Moreover, this thesis provides two original image data sets, which are made publicly available for the research community: the Spectral Image Data set of Religious Icon paintings (SIDRI) and the Spectral Texture data set of textile samples (SpecTex).
SIDRI is a multi-camera data set, taken with several RGB, multispectral, and hyperspectral cameras, which are spatially registered. The registration process makes this data set ideal for the testing and comparison of estimation models and camera design analysis.
SpecTex is one of very few available hyperspectral texture data sets that is used in this work to test the proposed feature extraction method in texture analysis application. These data sets are provided in a novel data structure based on the Tagged Image File Format (TIFF) that improves the flexibility and the user experience for viewing and archiving spectral image cubes as examples of big data.
The doctoral dissertation of Arash Mirhashemi, Master of Science, entitled Hyperspectral image acquisition, estimation, and feature extraction of complex and textured surfaces, will be publicly examined online, in English, on the 5th of August 2020, starting at 11 o'clock, at the Faculty of Science and Forestry. The opponent in the public examination will be Professor Masahiro Yamaguchi, Tokyo Institute of Technology, Japan, and the custos will be Professor Markku Hauta-Kasari, University of Eastern Finland.