These are preliminary reports that have not been peer-reviewed. They should not be regarded as conclusive, guide clinical practice/health-related behavior, or be reported in news media as established information. For more information, please see our FAQs.
LC_Paper_Final_v3.pdf (2.82 MB)
Convolutional Network Analysis of Optical Micrographs
Preprints are manuscripts made publicly available before they have been submitted for formal peer review and publication. They might contain new research findings or data. Preprints can be a draft or final version of an author's research but must not have been accepted for publication at the time of submission.
revised on 02.03.2020 and posted on 03.03.2020by Alexander Smith, Nicholas L. Abbott, Victor M. Zavala
We provide an in-depth convolutional neural network (CNN) analysis of optical responses of
liquid crystals (LCs) when exposed to different chemical environments. Our aim is to identify
informative features that can be used to construct automated LC-based chemical sensors and that
can shed some light on the underlying phenomena that governs and distinguishes LC responses.
Previous work demonstrated that, by using features extracted from AlexNet, micrographs of
different LC responses can be classified with an accuracy of 99%. Reaching such high levels of
accuracy, however, required use of a large number of features (on the order of thousands), which
was computationally intensive and which clouded the physical interpretability of the dominant
features. To address these issues, here we report a study of the effectiveness of using features
extracted from color images using VGG16, which is a more compact CNN than Alexnet. Our
analysis reveals that features extracted from the first and second convolutional layers of VGG16
are sufficient to achieve a perfect classification accuracy on the same dataset used by Cao and coworkers, while reducing the number of features to less than a hundred. The number of features
is further reduced to ten via recursive feature elimination with minimal loss in classification accuracy (5-10%). This feature reduction procedure reveals that differences in spatial color patterns are
developed within seconds in the LC response. The results thus reveal that hue histograms provide an informative set of features that can be used to characterize LC micrographs of the sensor
response. We also hypothesize that differences in spatial correlation length of LC textures detected
by VGG16 with DMMP and water likely reflect differences in the anchoring energy of the LC on the
surface of the sensor. This latter proposal hints at fresh approaches for the design of LC-based sensors based on characterization of spontaneous fluctuations in orientation (as opposed to changes
in time-average orientation)