Lens-free Multi-Laser Spectral Light-Field Fusion Microscopy

<p>We present a device and method for performing lens-free spectral<br />light-field fusion microscopy at sub-pixel resolutions while taking<br />advantage of the large field-of-view capability. A collection of<br />lasers at different wavelengths is used in pulsed mode and enables<br />the capture of interferometric light-field encodings of a specimen<br />placed near the detector. Numerically fusing the spectral complex<br />light-fields obtained from the encodings produces an image of the<br />specimen at higher resolution and signal-to-noise-ratio while suppressing<br />various aberrations and artifacts.</p>

A Discretize-then-Optimize Approach to Super-Resolution Reconstruction and Motion Estimation

<p>The process of recovering a high-resolution (HR) image from a set<br />of distorted (i.e. deformed, blurry, noisy, etc.) low-resolution (LR) images<br />is known as super-resolution. Super-resolution problem will require<br />the reconstruction of the HR image and estimations of motion<br />between LR images. In this study, image reconstruction and motion<br />estimation will be treated as a coupled problem. The proposed algorithm<br />uses an inverse model followed by a discretize-then-optimize<br />approach. Preliminary experiments on test data will be presented.</p>

A Conditional Random Field Weakly Supervised Segmentation Approach for Segmenting keratocytes Cells in Corneal Optical Coherence Tomography Images

<p>Keratocytes are vital for maintaining the overall health of human<br />cornea as they preserve the corneal transparency and help in healing<br />corneal injuries. Manual segmentation of keratocytes is challenging,<br />time consuming and also needs an expert. Here, we propose<br />a novel semi-automatic segmentation framework, called Conditional<br />Random FieldWeakly Supervised Segmentation (CRF-WSS)<br />to perform the keratocytes cell segmentation. The proposed framework<br />exploits the concept of dictionary learning in a sparse model<br />along with the Conditional Random Field (CRF) modeling to segment<br />keratocytes cells in Ultra High Resolution Optical Coherence<br />Tomography (UHR-OCT) images of human cornea. The results<br />show higher accuracy for the proposed CRF-WSS framework compare<br />to the other tested Supervised Segmentation (SS) andWeakly<br />Supervised Segmentation (WSS) methods.</p>

Efficient and Scalable Image Segmentation Using Bag-of-Features and Stochastic Region Merging

<p>This work presents an efficient and scalable texture segmentation<br />algorithm based on bag-of-features and stochastic region merging.<br />The image is partitioned into blocks and processed independently<br />to obtain regions, which are then merged to obtain the final<br />segmentation. Experimental results shows the proposed method<br />achieves an overall speed improvement of at least 4.5x and requires<br />6.5x less memory, while still improving segmentation accuracy<br />for large images.</p>

Colour-based gesture recognition for American Sign Language via Hidden Markov Models

<p>We present a new approach to gesture recognition for use in a sign<br />language learning environment. This method utilizes inexpensive<br />cloth gloves to alleviate the difficulty of hand detection and to allow<br />for feature creation. Salient colours identify the glove base and<br />fingertip markers, which are then used to extract a hand centroid<br />and a convex hull describing the fingertips for each hand. A Hidden<br />Markov Model is created for each sign, as well as an additional<br />threshold model created from all signs. When a candidate sign is<br />performed, the sign of the HMM that produces the greatest likelihood<br />is matched, provided it also exceeds the threshold model<br />likelihood. Isolated recognition testing of the training library indicated<br />76% accuracy, and continuous recognition testing showed<br />60% accuracy.</p>

Structural Representation: Reducing Multi-Modal Image Registration to Mono-Modal Problem

<p>Registration of multi-modal images has been a challenging task<br />due to the complex intensity relationship between images. The<br />standard multi-modal approach tends to use sophisticated similarity<br />measures, such as mutual information, to assess the accuracy<br />of the alignment. Employing such measures imply the increase in<br />the computational time and complexity, and makes it highly difficult<br />for the optimization process to converge. The presented registration<br />method works based on structural representations of images<br />captured from different modalities, in order to convert the multimodal<br />problem into a mono-modal one. Two different representation<br />methods are presented. One is based on a combination of<br />phase congruency and gradient information of the input images,<br />and the other utilizes a modified version of entropy images in a<br />patch-based manner. Sample results are illustrated based on experiments<br />performed on brain images from different modalities.</p>

A Portable Plug-and-Play Imaging System for Physiological Monitoring

<p>Non-contact physiological monitoring has garnered interest in the<br />research community. However, studies employ different imaging<br />system configuration, including illuminant profile and wavelength,<br />camera type and frame rate, and distance to the skin. In this paper,<br />we propose an easily customizable imaging system for evaluating<br />physiological monitoring parameters. The system’s design allows<br />plug-and-play compatibility with various illumination sources,<br />camera types, lenses, and optical components. Results using one<br />configuration shows the feasibility of spatial blood perfusion analysis<br />using this imaging system, where areas exhibiting clean blood<br />pulsatility can be identified at the surface level.</p>

Stochastic Receptive Fields in Deep Convolutional Networks

<p>Deep convolutional neural networks (ConvNets) have rapidly grown<br />in popularity due to their powerful capabilities in representing and<br />modelling the high-level abstraction of complex data. However,<br />ConvNets require an abundance of data to adequately train network<br />parameters. To tackle this problem, we introduce the concept<br />of stochastic receptive fields, where the receptive fields are<br />stochastic realizations of a random field that obey a learned distribution.<br />We study the efficacy of incorporating layers of stochastic<br />receptive fields to a ConvNet to boost performance without the<br />need for additional training data. Preliminary results showing an<br />improvement in accuracy ( 2% drop in test error) was achieved by<br />adding a layer of stochastic receptive fields to a ConvNet compared<br />to adding a layer of fully-trained receptive fields, when training with<br />a small training set consisting of 20% of the STL-10 dataset.</p>

Using Superpixels for Road Segmentation in Street View Images

Google Street View is a useful database that houses a large amount of information. This information, however, is unlabelled. We explore the use of superpixel methods for segmentation of images in this database, specifically road segmentation.

An Integrated Systems Design Framework for Computational Polarimetry

<p>In this study, we propose and implement an integrated systems<br />design framework for computational polarimetry. This framework<br />leverages knowledge of the optical elements to aid in the design of<br />the polarimetry instrumentation and the enhancement of the measurements.<br />This framework incorporates the use of spatial detector<br />arrays, and models the non-ideal performance of the optical components,<br />providing error bounds that can decrease the cost of the<br />system depending on the accuracy needed. Noise modelling is incorporated,<br />as well, in the measurement formation models. The<br />framework is demonstrated in the design of a computational polarimetry<br />system for a glucose concentration estimation.</p>