A Deep Neural Network for Fast Confocal Laser Scanning Microscopy Imaging Recovery Algorithm from Random Undersampling Measurements

• Main Authors: Kuang-Yao Chang, 張光耀
• Other Authors: 傅立成
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/avr4ue

碩士 === 國立臺灣大學 === 電機工程學研究所 === 105 === Confocal laser scanning microscopy (CLSM) is a powerful non-destructive optical inspection system in high precision measurement technology. CLSM can be used to construct three-dimensional profile of biological cells or micro and sub-micro engineering materials. Recently, compressive sensing (CS) is applied to CLSM system for high speed scan by reducing the amount of sampled data required to reconstruct an accurate image. However, the CS recovery algorithm employed in CLSM applications is iteration-based optimization method of which computation complexity is relatively high. In this work, a non-iteration-based deep residual convolutional neural network compressive sensing reconstruction (DRCNN-CSR) framework in end-to-end manner is proposed. Not only the computation time but also the quality of reconstructed image is greatly improved with this algorithm, and our method has an ability to recover images sampled under multi-undersampling rates (USRs) in single trained model. The quantitative comparisons with state-of-the-art CS recovery algorithms are provided, and the experiment results demonstrate that our proposed method outperforms the others under a wide range of under-sampling rates. Furthermore, in order to deal with the uneven sample information density problem, we also propose an adaptive undersampling rate strategy to adjust the under-sampling rate in different local areas. At the end, by the reconstructions of the real CLSM measured data in random scanning pattern, the recovery robustness of our model is validated for fast CLSM imaging application.