Abstract
Hyperspectral imaging (HSI) has become an essential tool for exploration of different spatially-resolved properties of materials in analytical chemistry. However, due to various technical factors such as detector sensitivity, choice of light source and experimental conditions, the recorded data contain noise. The presence of noise in the data limits the potential of different data processing tasks such as classification and can even make them ineffective. Therefore, reduction/removal of noise from the data is a useful step to improve the data modelling. In the present work, the potential of a wavelength-specific shearlet-based image noise reduction method was utilised for automatic de-noising of close-range HS images. The shearlet transform is a special type of composite wavelet transform that utilises the shearing properties of the images. The method first utilises the spectral correlation between wavelengths to distinguish between levels of noise present in different image planes of the data cube. Based on the level of noise present, the method adapts the use of the 2-D non-subsampled shearlet transform (NSST) coefficients obtained from each image plane to perform the spatial and spectral de-noising. Furthermore, the method was compared with two commonly used pixel-based spectral de-noising techniques, Savitzky-Golay (SAVGOL) smoothing and median filtering. The methods were compared using simulated data, with Gaussian and Gaussian and spike noise added, and real HSI data. As an application, the methods were tested to determine the efficacy of a visible-near infrared (VNIR) HSI camera to perform non-destructive automatic classification of six commercial tea products. De-noising with the shearlet-based method resulted in a visual improvement in the quality of the noisy image planes and the spectra of simulated and real HSI. The spectral correlation was highest with the shearlet-based method. The peak signal-to-noise ratio (PSNR) obtained using the shearlet-based method was higher than that for SAVGOL smoothing and median filtering. There was a clear improvement in the classification accuracy of the SVM models for both the simulated and real HSI data that had been de-noised using the shearlet-based method. The method presented is a promising technique for automatic de-noising of close-range HS images, especially when the amount of noise present is high and in consecutive wavelengths.
Highlights
Close-range hyperspectral imaging (HSI) and image processing techniques are popular analytical tools in many scientific domains and are used in applications such as the exploration of food properties [1], pharmaceutical product characterisation [2,3], forensics analysis [4,5], exploration of plant traits for phenotype studies [6,7], and microbiology [8]
De-noising and classification experiments were performed with visible-near infrared (VNIR) hyperspectral images of six different commercial tea products, which were purchased from a local market (Glasgow, United Kingdom)
Used methods such as SAVGOL smoothing and median filtering can deal with a small amount of noise and if the noise is not present in neighbouring wavelengths
Summary
Close-range hyperspectral imaging (HSI) and image processing techniques are popular analytical tools in many scientific domains and are used in applications such as the exploration of food properties [1], pharmaceutical product characterisation [2,3], forensics analysis [4,5], exploration of plant traits for phenotype studies [6,7], and microbiology [8]. HSI combines two sensor modalities that are spectroscopy and imaging, where the spectroscopy provides the chemical information about the samples and the imaging adds a complementary domain of spatial information [9]. The data generated by HSI can be understood as spatial maps of spectral variation arranged in 3-D cubes (n × p × q). The first two dimensions (n × p) of the cubes are usually the spatial dimensions, and the third dimension (q) contains the spectral information. Chemical 281 (2019) 1034–1044 different data processing steps such as exploration, regression and classification are often performed. Before any data processing, as a standard first step, the cubes are usually pre-processed to remove various types of noise from the data so as to increase the signalto-noise ratio (SNR) [10]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
