Design considerations for line-scan multi-spectral imaging systems

Resumen

Spectral imaging systems have been used for spectral measurements for several decades, mainly for scientific purposes, and were usually linked to costly applications. Thanks to research in imaging science and machine learning and the technological advancement in recent years, spectral imaging became feasible for various industrial and even consumer applications. This dissertation deals with line-scan multi-spectral imaging systems for spectral reflectance and color measurements. The most important aspects under consideration are optimization of spectral properties of the optical components, image registration and estimation of surface spectral reflectances. We focus on a particular system design, in which multiple color filtered RGB images with distinct spectral content are acquired at the same time. These images correspond to different viewpoints of the scanning scene due to the mechanical arrangement of camera sensor and optics. By optimizing the system's optical component spectral properties, the amount of spectral information acquired can be increased and the spectral reflectance estimation can be improved. We propose a filter selection framework and demonstrate that optimization for various line-scan system configurations results in an improvement of spectral and color measurement performance. Multi-channel image registration is required to account for viewpoint differences and other sources of image channel misalignment. We develop a calibration scheme for planar scanning objects and propose scene-adaptive registration for non-planar scanning objects. For our 12-channel laboratory imaging system, sub-pixel accuracy is achieved. Based on the registered multi-channel image data, spectral reflectance estimation can be performed. Physical and empirical estimation methods are considered, and we propose a logarithmic kernel function for kernel ridge regression. We experimentally compare performance of various estimation methods for simulated and measured camera response data and consider different noise levels and number of spectral channels. Empirical estimation performance is influenced by model training. We compare various training sample selection approaches and propose an application dependent selection scheme. Further, adaptive training methods from related literature are unified conceptually and evaluated systematically. We show that the aforementioned aspects of line-scan multi-spectral imaging system design are critical for spectral and color measurement, and that application specific design is often beneficial to improve system performance.