2018/10/19(五) 14:20 - 林家祥教授(中央大學太空及遙測研究中心系) - Hyperspectral Unmixing and Super-resolution (高光譜解混與超解析)
Topic:
Hyperspectral Unmixing and Super-resolution (高光譜解混與超解析)
Speaker:
Prof. 林家祥 Chia-Hsiang Lin
中央大學太空及遙測研究中心
Center for Space and Remote Sensing Research,NCU
Language :
Chinese
Date:
2018/10/19
Time:
14:20~16:20
Location: Delta building R216, NTHU
Abstract:
Hyperspectral unmixing (HU), aiming at unsupervisedly recovering the underlying spectral signatures of source materials, has found many successful scientific applications, including remote sensing and bioinformatics. HU also has strong connections to problems from related fields such as machine learning, data analytics and analytical chemistry. Unlike conventional HU methods, convex geometry (CG) based HU does not require unrealistic statistical assumptions (e.g., source independence). A well-known CG criterion is rooted in the so-called Craig simplex, which has led to substantial advances in HU in the past two decades. Nevertheless, Craig estimator induces a non-convex optimization problem, and such non-convexity seriously degrades its effectiveness when the sources are heavily mixed, as demonstrated in our recent SIAM paper. As a breakthrough, we recently employed the John ellipsoid, a seminal concept in functional analysis, to design a brand new criterion, for which we proved that its source identifiability is as strong as that of Craig simplex, but only requiring solving a conic programming that is convex. Moreover, we surprisingly found that John ellipsoid also creates a new door to attack the tough source separation scenario wherein the mixing system is ill-conditioned (an issue bothering biologists for years). On the other hand, to mitigate the Poisson noise, spatial resolution of hyperspectral image (HSI) is often limited; this is the case especially for remotely acquired satellite imagery. Without relying on hardware improvement, super-resolving the HSI computationally is fundamental and essential for the subsequent computer vision tasks. Nevertheless, hundreds of densely populated spectral bands in typical HSI make such hyperspectral super-resolution (HySure) problem highly challenging. In this talk, an image fusion based HySure method, which achieves state-of-the-art performance on four key indices and over three hyperspectral sensors, will also be presented.
Biography:
清華大學電機學士、通訊博士。
任職於中央大學前，於香港中文大學、Virginia Tech、里斯本大學，進行跨領域研究。
博士班期間於 Virginia Tech 參與美國國立衛生研究院贊助的計畫，研究基因表達、生醫影像。
博士後期間於里斯本大學參與歐洲太空總署贊助的計畫，研究衛星遙測、高光譜影像。
博士論文獲得 IEEE 地科及遙測學會 Best Thesis Award、中華民國影像處理與圖形識別學會 Outstanding Thesis Award。
自 Aug. 2018 至 July 2023 五年期間，為科技部愛因斯坦培植計畫的主持人。
Hyperspectral Unmixing and Super-resolution (高光譜解混與超解析)
Speaker:
Prof. 林家祥 Chia-Hsiang Lin
中央大學太空及遙測研究中心
Center for Space and Remote Sensing Research,NCU
Language :
Chinese
Date:
2018/10/19
Time:
14:20~16:20
Location: Delta building R216, NTHU
Abstract:
Hyperspectral unmixing (HU), aiming at unsupervisedly recovering the underlying spectral signatures of source materials, has found many successful scientific applications, including remote sensing and bioinformatics. HU also has strong connections to problems from related fields such as machine learning, data analytics and analytical chemistry. Unlike conventional HU methods, convex geometry (CG) based HU does not require unrealistic statistical assumptions (e.g., source independence). A well-known CG criterion is rooted in the so-called Craig simplex, which has led to substantial advances in HU in the past two decades. Nevertheless, Craig estimator induces a non-convex optimization problem, and such non-convexity seriously degrades its effectiveness when the sources are heavily mixed, as demonstrated in our recent SIAM paper. As a breakthrough, we recently employed the John ellipsoid, a seminal concept in functional analysis, to design a brand new criterion, for which we proved that its source identifiability is as strong as that of Craig simplex, but only requiring solving a conic programming that is convex. Moreover, we surprisingly found that John ellipsoid also creates a new door to attack the tough source separation scenario wherein the mixing system is ill-conditioned (an issue bothering biologists for years). On the other hand, to mitigate the Poisson noise, spatial resolution of hyperspectral image (HSI) is often limited; this is the case especially for remotely acquired satellite imagery. Without relying on hardware improvement, super-resolving the HSI computationally is fundamental and essential for the subsequent computer vision tasks. Nevertheless, hundreds of densely populated spectral bands in typical HSI make such hyperspectral super-resolution (HySure) problem highly challenging. In this talk, an image fusion based HySure method, which achieves state-of-the-art performance on four key indices and over three hyperspectral sensors, will also be presented.
Biography:
清華大學電機學士、通訊博士。
任職於中央大學前，於香港中文大學、Virginia Tech、里斯本大學，進行跨領域研究。
博士班期間於 Virginia Tech 參與美國國立衛生研究院贊助的計畫，研究基因表達、生醫影像。
博士後期間於里斯本大學參與歐洲太空總署贊助的計畫，研究衛星遙測、高光譜影像。
博士論文獲得 IEEE 地科及遙測學會 Best Thesis Award、中華民國影像處理與圖形識別學會 Outstanding Thesis Award。
自 Aug. 2018 至 July 2023 五年期間，為科技部愛因斯坦培植計畫的主持人。