Author:

Abstract:

Hyperspectral imaging increases the capabilities of traditional machine vision by extending the information content from three broad bands (RGB) to a spectrum of multiple narrow bands beyond the visible domain. This provides a combination of spectral and spatial information, which increases the potential for applications with respect to traditional color imaging or point spectroscopy. While hyperspectral imaging is a technology that has already shown high potential in a wide range of application domains, its adoption by Industry has been slow so far. This has been attributed to the high camera cost on one hand and processing expertise required for the large amounts of data generated on the other hand. In this sense, recent hyperspectral technology developments are trying to bridge this gap by creating more affordable cameras that can better meet industrial needs. Typically, the development of more industrially suited cameras is done at the expense of either a lower number of bands or lower spatial resolution, which may in turn reduce their discrimination performance with respect to high-end research equipment. To explore these trade-offs, a system-wide exploration was performed of hyperspectral imaging based on cameras, which target industrial needs. To this end, multiple system parameters such as wavelength range, camera hardware, illumination system or data analysis methods were varied for some specific applications. First, system level optimization was explored by using the wavelength range as a key system parameter to reduce camera hardware cost for a textile sorting application. In this application, it is shown that a suboptimal wavelength range may still be able to meet the discrimination requirements, while substantially reducing the hardware cost. Next, the focus was shifted to a case of seed mix ingredient discrimination and quantification. The added value of data preprocessing and the integration of spatial information with the spectral information is demonstrated to increase the system performance and reach the application targets. Further, it is demonstrated that the illumination system is a key parameter in hyperspectral imaging applications, in particular with snapshot cameras. The presented results show how illumination can have a relevant impact on the performance (up to 10% increase in classification accuracy) by achieving a more balanced spectral and spatial illumination. Finally, different system parameters such as camera hardware, illumination system and data analysis methods are evaluated together. In terms of data processing, the impact of pre- and post-processing methods are explored, while pixel-based analysis is compared to a more joint spatial-spectral image analysis based on convolutional neural networks. It is demonstrated that the joint evaluation of all these system parameters allows to make the best choices to meet the application requirements and increased the mean classification accuracy by up to 25%. Moreover, it allows to explore varied system configurations that offer different performance-cost-speed tradeoffs. To conclude this dissertation, some guidelines for system level optimization and parameter selection are proposed from the application characteristics and requirements. This paves the way for a broader industrial adoption of hyperspectral imaging technology.