Experimental demonstration of an adaptive architecture for direct spectral imaging classification

Dunlop-Gray, Matthew; Poon, Phillip K.; Golish, D. R.; Vera, Esteban; Gehm, Michael E.

doi:10.1364/oe.24.018307

Cited by 14 publications

(7 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…First, our optimized LED patterns require two images (assuming positive and negative weights). Following recent work with spectral classification [32], a future strategy may use an adaptive approach to classify with more than two uniquely lit images (e.g., the LED pattern for each subsequent image can depend on the classification scores for previous images). Second, there are a large number of alternative physical aspects of a microscope (or an imaging system in general) to potentially optimize via a CNN.…”

Section: Resultsmentioning

confidence: 99%

Convolutional neural networks that teach microscopes how to image

Horstmeyer,

Chen,

Kappes

et al. 2017

Preprint

View full text Add to dashboard Cite

Deep learning algorithms offer a powerful means to automatically analyze the content of medical images. However, many biological samples of interest are primarily transparent to visible light and contain features that are difficult to resolve with a standard optical microscope. Here, we use a convolutional neural network (CNN) not only to classify images, but also to optimize the physical layout of the imaging device itself. We increase the classification accuracy of a microscope's recorded images by merging an optical model of image formation into the pipeline of a CNN. The resulting network simultaneously determines an ideal illumination arrangement to highlight important sample features during image acquisition, along with a set of convolutional weights to classify the detected images post-capture. We demonstrate our joint optimization technique with an experimental microscope configuration that automatically identifies malaria-infected cells with 5-10% higher accuracy than standard and alternative microscope lighting designs.

show abstract

Section: Resultsmentioning

confidence: 99%

Convolutional neural networks that teach microscopes how to image

Horstmeyer,

Chen,

Kappes

et al. 2017

Preprint

View full text Add to dashboard Cite

show abstract

“…The obvious drawback with this method is the loss in spatial resolution, albeit in one direction only, as we only need to consider the direction of dispersion -along the horizontal axis. Pixel binning has been used in some recent works, [19,17], although with little discussion of the advantages or disadvantages.…”

Section: Pixel Binning and Averagingmentioning

confidence: 99%

“…There is also interest in reducing the number of acquisitions required per datacube; recent work has measured the spectrum using a Fourier basis, which can reduce the number of acquisitions, as long as there are no sharp spectral features [17]. Classification of multiple spectra is possible with only a few acquisitions as shown in [18,19], using a scheme where the DMD mask adapts for each new acquisition to improve classification accuracy. Furthermore, there is potential to use regularization approaches to exploit homogeneity within the hyperspectral datacube, allowing full cube reconstruction with only a handful of acquisitions [20].…”

Section: Introductionmentioning

confidence: 99%

Calibration of programmable spectral imager with dual disperser architecture

Hemsley

Lacroix

Carfantan

et al. 2020

Optics Communications

View full text Add to dashboard Cite

We present a hyperspectral imager with a dual disperser architecture, which is highly re-configurable thanks to the use of a digital micro-mirror device (DMD). The system has over 200 spectral bands in the range 400-700 nm, and 2048 by 2048 spatial pixels. We describe a method to calibrate the instrument and quantify its performance using a simple slit scanning acquisition scheme, and discuss various methods to remove artifacts in the spectral datacube via postprocessing. This prototype instrument is designed for simplicity, and aims to explore new acquisition schemes that exploit the property of colocation and use the programmable nature of the DMD to increase the speed of datacube retrieval and the quality of the hyperspectral data.

show abstract

“…This method is severely affected by classification noise and it is computationally demanding. A classification method in the compressive domain is also reported in [27] that is relied on an adaptive sensing scheme [22]. This approach consists of an adaptive coded aperture design method based on the Bayesian classifier.…”

Section: Introductionmentioning

confidence: 99%

Adaptive Multisensor Acquisition via Spatial Contextual Information for Compressive Spectral Image Classification

Díaz¹,

Ramirez²,

Vera³

et al. 2021

IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing

Self Cite

View full text Add to dashboard Cite

Spectral image classification uses the huge amount of information provided by spectral images to identify objects in the scene of interest. In this sense, spectral images typically contain redundant information that is removed in later processing stages. To overcome this drawback, compressive spectral imaging (CSI) has emerged as an alternative acquisition approach that captures the relevant information using a reduced number of measurements. Various methods that classify spectral images from compressive projections have been recently reported whose measurements are captured by non-adaptive, or adaptive schemes discarding any contextual information that may help to reduce the number of captured projections. In this paper, an adaptive compressive acquisition method for spectral image classification is proposed. In particular, we adaptively design coded aperture patterns for a dual-arm CSI acquisition architecture, where the first system obtains compressive multispectral projections and the second arm registers compressive hyperspectral snapshots. The proposed approach exploits the spatial contextual information captured by the multispectral arm to design the coding patterns such that subsequent snapshots acquire the scene's complementary information improving the classification performance. Results of extensive simulations are shown for two state-of-theart databases: Pavia University and Indian Pines. Furthermore, an experimental setup that performs the adaptive sensing was built to test the performance of the proposed approach on a real dataset. The proposed approach exhibits superior performance with respect to other methods that classify spectral images from compressive measurements.

show abstract

Experimental demonstration of an adaptive architecture for direct spectral imaging classification

Cited by 14 publications

References 24 publications

Convolutional neural networks that teach microscopes how to image

Convolutional neural networks that teach microscopes how to image

Calibration of programmable spectral imager with dual disperser architecture

Adaptive Multisensor Acquisition via Spatial Contextual Information for Compressive Spectral Image Classification

Contact Info

Product

Resources

About