Ultrahigh-resolution spectral CCD imaging based on machine learning

• Main Author: Xie, Hui
• Other Authors: Y. C. Chen
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Optics, optoelectronics, photonics
• Online Access: https://hdl.handle.net/10356/153855
• Institution: Nanyang Technological University

Spectral imaging has played a significant role in many applications, including earth remote sensing, military, and biomedical solutions. The request for higher spectral resolution has been in demand, especially for laser emission-based detection and imaging. Various systems and configurations have been developed over the past decades. However, state-of-art cameras relies heavily on optical filters or a bulky spectrometer system to achieve high spectral resolution. As such, the ability to obtain wavelength information from normal CCD images remains a key challenge. Herein, we developed a smart algorithm to identify emission wavelengths captured by a conventional CCD camera. A total of 6000 CCD images with precise laser emissions from 400 nm to 670 nm were used to train the model. Red, green, blue channels and its corresponding grey value were extracted as the pixel feature. Three advanced modeling methods were employed based on the Python library scikit-learn, namely, decision tree, k-nearest neighbor (KNN) and gradient boosting regression (GBR) model. Under the same wavelength and luminance condition, the reranked feature between the tenth and ninetieth percentile remained unchanged, otherwise modified to the median. Based on the model developed, we applied the algorithm to a multi-wavelength image captured from CCD for real imaging applications of microparticles. Our results showed the possibility of spectral mapping with a high spectral resolution of 0.98 nm while maintaining pixel-sized spatial resolution. The proposed method could be applied to a wide range of imaging applications.