Eric D. Sinzinger

A model-based approach to junction detection using radial energy

A novel junction detector is presented that fits the neighborhood around a point to a junction model. The junction model segments the neighborhood into wedges by determining a set of radial edges. The radial edges are invariant to affine transforms, creating an affine invariant junction detector. The radial edges are evaluated based upon the pixels along the edge. The angle between the pixel gradient and the vector to the potential junction point forms the initial basis for the measurement. An initial set of radial edges is selected based upon identifying local maximums within a given arc distance. An energy function is applied to the resulting radial segmentation, and a greedy optimization routine is used to construct the minimal set of radial edges. To identify the final junctions, a second energy function is used that combines the components of the first energy function with the resulting change in standard deviation by separation into radial segments. The junctions with the most energy in their local neighborhoods are selected as potential junctions. The neighborhoods about the potential junctions are analyzed to determine if they represent a single line or multiple non-parallel lines. If the neighborhood represents multiple non-parallel lines, the point is classified as a junction point. The junction detector is tested on several images including both synthetic and real images. Highlights of radially segmented junction points are displayed for the real images.

Improvements in the predictive capability of neural networks

Abstract Neural networks can be used to develop effective models of nonlinear systems. Their main advantage being that they can model the vast majority of nonlinear systems to any arbitrary degree of accuracy. The ability of a neural network to predict the behavior of a nonlinear system accurately ought to be improved if there was some mechanism that allows the incorporation of first-principles model information into their training. This study proposes to use information obtained from a first-principle model to impart a sense of “direction” to the neural network model estimate. This is accomplished by modifying the objective function so as to include an additional term that is the difference between the time derivative of the outputs, as predicted by the neural network, and that of the outputs of the first-principles model during the training phase. The performance of a feedforward neural network model that uses this modified objective function is demonstrated on a chaotic process and compared to the conventional feedforward network trained on the usual objective function.

Classification of Cervix Lesions Using Filter Bank-Based Texture Mode

This paper explores the classification of texture patterns observed in digital images of the cervix. In particular, the problem of identifying and segmenting punctations and mosaic patterns is considered. First, the ability of large scale filter banks in characterizing punctations and mosaic structures is studied using texton models. However, texton-based models fail to consistently classify punctation and mosaic sections obtained from cervix images of different subjects. We present a novel method to segment punctations that combines matched filtering using a Gaussian template with Gaussian mixture models. Features extracted from the objects detected using this novel method on punctation and mosaic sections are shown to provide excellent classification between punctation and mosaicism. Results demonstrate the effectiveness of our approach in detecting punctations and separating punctation sections from mosaic sections

Next best view algorithms for interior and exterior model acquisition

Rapid determination of optimal laser range finder placement is becoming of increased importance in both computer vision and robotics. The need of next-best-view algorithms for intelligent decisions in where to place the laser range finder is important for scanning both objects and landscapes to assure that the scene is fully realized and can be registered accurately. Presented here are methods for determining sensor placement to maximize the amount of information that is learned about a scene or object. Using occupancy grids of voxels and ray tracing, the placement of the sensor can be optimized for maximum collection of new data. This work provides fast algorithms that determine optimal sensor placement both outside an object of interest and inside of a closed environment. These methods take into consideration the limitations of the sensor, the interaction of the sensor to the environment, and its placement in only known areas to restrict the search space.

Image Matching Using High Dynamic Range Images and Radial Feature Descriptors

Obtaining a top match for a given query image from a set of images forms an important part of the scene identification process. The query image typically is not identical to the images in the data set, with possible variations of changes in scale, viewing angle and lighting conditions. Therefore, features which are used to describe each image should be invariant to these changes. Standard image capturing devices lose much of the color and lighting information due to encoding during image capture. This paper uses high dynamic range images to utilize all the details obtained at the time of capture for image matching. Once the high dynamic range images are obtained through the fusion of low dynamic range images, feature detection is performed on the query images as well as on the images in the database. A junction detector algorithm is used for detecting the features in the image. The features are described using the wedge descriptor which is modified to adapt to high dynamic range images. Once the features are described, a voting algorithm is used to identify a set of top matches for the query image.

SEDATION OF SIMULATED ICU PATIENTS USING REINFORCEMENT LEARNING BASED CONTROL

The Intensive Care Unit (ICU) is a challenging environment to both patient and caregiver. Continued shortages in staffing increase risk to patients. To evaluate the use of intelligent systems in the improvement of patient care, an intelligent agent was developed to regulate ICU patient sedation. A temporal differencing form of reinforcement learning was used to train the agent in the administration of intravenous propofol in simulated ICU patients. The agent utilized a well-studied pharmacokinetic model to calculate the distribution of drug within the patient. Pharmacodynamics were then estimated for the drug effect. A derivative of the electroencephalograms, the bispectral index, served as the system control variable. The agent demonstrated satisfactory control of the simulated patients consciousness level in static and dynamic setpoint conditions. The agent demonstrated superior stability and responsiveness when compared to a well-tuned PID controller, the control method of choice in closed-loop sedation control literature.

Scene Identification Using Invariant Radial Feature Descriptors

This paper addresses the challenge of identifying and retrieving related scenes from image databases with a focus on low-level feature descriptor construction. A set of affine covariant regions are identified via a radial segmentation algorithm. Local descriptors are built using two different types of histograms: (i) polar image gradient (PIG) orientation histogram, and (ii) saturation-weighted hue histogram. The combination of geometric and photometric information yields a significant improvement in a features discriminative power. A cascading matching algorithm is used for feature matching and evaluation. To demonstrate the descriptors image matching capabilities, a voting algorithm for similar scene retrieval is implemented utilizing results from the feature matches. Challenging images of buildings with inherent replicative feature regions due to common edificial texture are used to test the robustness and applicability of the radial-based methodology.

Autonomous map construction using three-dimensional feature descriptors

Autonomous robotic mapping has been an open research topic for more than twenty years. The primary objective of the robotic mapping problem is to design methods that can guide a robot around an environment and allow it to create a map of what has been sensed. Most automatic mapping algorithms rely on robot pose estimation to fuse map data together. This paper demonstrates that through feature extraction using spin-histograms, the pose of the robot can be estimated accurately enough for an Iterative Closest Point (ICP) algorithm to register overlapping data sets. By eliminating consideration for points according to curvature and saliency, the spin-histogram matching process can improve in both accuracy and computation time. In combination with a global registration algorithm known as simultaneous matching, this process can obtain a fully autonomous registration process.

Cascading Trilinear Tensors for Face Authentication

This paper presents a method to improve the accuracy rates of face authentication between images with different poses. Trilinear tensors are used to adjust the pose of the training and testing images. All the images are transformed by a pose adjustment algorithm so novel images are generated that have the same pose. These novel images are then used to train and test support vector machine (SVM) face authentication functions to verify the identity of the people in the images. The overall results show that the accuracy improves when the poses of the images are adjusted.