Josh Harguess

Image Classification Using Generative Neuro Evolution for Deep Learning

Research into deep learning has demonstrated performance competitive with humans on some visual tasks, however, these systems have been primarily trained through supervised and unsupervised learning algorithms. Alternatively, research is showing that evolution may have a significant role in the development of visual systems. Thus neuroevolution for deep learning is investigated in this paper. In particular, the Hypercube-based Neuro Evolution of Augmenting Topologies is a NE approach that can effectively learn large neural structures by training an indirect encoding that compresses the artificial neural network (ANN) weight pattern as a function of geometry. The methodologies are tested on a traditional image classification task as well as one tailored to overhead satellite imagery. The results show that Hyper NEAT struggles with performing image classification by itself, but can be effective in training a feature extractor that other ML approaches can learn from. Thus Neuro Evolution combined with other ML methods provides an intriguing area of research that can replicate the processes in nature.

Feature Learning HyperNEAT: Evolving Neural Networks to Extract Features for Classification of Maritime Satellite Imagery

Imagery analysis represents a significant aspect of maritime domain awareness; however, the amount of imagery is exceeding human capability to process. Unfortunately, the maritime domain presents unique challenges for machine learning to automate such analysis. Indeed, when object recognition algorithms observe real-world data, they face hurdles not present in experimental situations. Imagery from such domains suffers from degradation, have limited examples, and vary greatly in format. These limitations are present satellite imagery because of the associated constraints in expense and capability. To this end, the Hypercube-based NeuroEvolution of Augmenting Topologies approach is investigated in addressing some such challenges for classifying maritime vessels from satellite imagery. Results show that HyperNEAT learns features from such imagery that allows better classification than Principal Component Analysis (PCA). Furthermore, HyperNEAT enables a unique capability to scale image sizes through the indirect encoding.

Structured outlier models for robust dictionary learning

Robust dictionary learning algorithms seek to learn a dictionary while being robust to the presence of outliers in the training set. Often, the elements of the training set have an underlying structure due to, for example, their spatial relation or their similarity. When outliers are present as elements of the training set, they often inherit the underlying structure of the training set. This work capitalizes on such structure, encoded as an undirected graph connecting elements of the training set, and on sparsity-aware outlier modeling tools to develop robust dictionary learning algorithms. Not only do these algorithms yield a robust dictionary, but they also identify the outliers in the training set. Computationally efficient algorithms based on block coordinate descent and proximal gradient methods are developed.

Online robust dictionary learning with density-based outlier weighing

Online outlier detection is fundamental for expediting the processing of data and focusing processing resources on portions of data that may be most informative. This work develops online robust dictionary learning algorithms that are able to identify outliers in the training data. The algorithms are based on lasso updates for computing the vector of expansion coefficients for a new training vector and gradient descent updates for updating the dictionary. An outlier is identified based on the so-called outlier vector. The weight associated with the group lasso regularizer that encourages an outlier vector to be set to zero is computed based on the outlierness score of the corresponding training data vector. Outlier vectors are thus more likely to be nonzero if they feature a high outlierness score. Outlierness scores are obtained from density-based outlier detection algorithms and help to enhance the selection of outliers. Both soft and hard outlier removal algorithms are developed. In the latter case, outliers are identified and a residual obtained after removing the outlier contribution is used to update the dictionary. The performance of the proposed algorithms is illustrated via numerical experiments on real video data.

Optical flow on degraded imagery

Motion estimation from video is an increasingly important problem with applications in ego-motion estimation of an unmanned vehicle, segmentation from video, object detection and tracking, and many others. Recent advances in optical flow have made motion estimation possible in many applications with high-resolution imagery. However, in the presence of noise and compression artifacts, these state-of-the-art optical flow algorithms fail, in various ways, to recover the motion where they would otherwise succeed. We study the effects of adding noise and compression artifacts to a well-known optical flow dataset when processed by several state-of-the-art algorithms and present compelling qualitative and quantitative results due to these degradations.

Classifying Maritime Vessels from Satellite Imagery with HyperNEAT

Maritime data uniquely challenges imagery analysis. Such data suffers from degradation, limited samples, and varied formats. To this end, the Hypercube-based NeuroEvolution of Augmenting Topologies (HyperNEAT) approach is investigated in addressing such challenges for classifying maritime vessels in a satellite imagery data set. The results show that HyperNEAT learns to extract features that allows better classification than those from Principal Component Analysis (PCA) and robust to differences in presentation of data. Furthermore, HyperNEAT enables a unique capability to scale trained solutions to different image resolutions.

Polynomial expansion for range image segmentation and classification of the environment

In this paper we introduce a method that utilizes a high-order polynomial expansion of range imagery for the purposes of image segmentation and classification. The use of polynomial expansion has been quite successful in segmenting and estimating optical flow in 2D imagery, but has not been used extensively in 3D or range imagery. We derive features using the coefficients of the high-order polynomial expansion and use those features for local and global segmentation of the range image. Finally, we classify the segments based on the features within each segment. Promising results are shown on range images from the Odetic lidar database.

Ego-Motion Estimation on Range Images Using High-Order Polynomial Expansion

This paper presents two novel algorithms for estimating the (local and global) motion in a series of range images based on a polynomial expansion. The use of polynomial expansion has been quite successful in estimating optical flow in 2D imagery, but has not been used extensively in 3D or range imagery. In both methods, each range image is approximated by applying a high-order polynomial expansion to local neighborhoods within the range image. In the local motion algorithm, these approximations are then used to derive the translation or displacement estimation within the local neighborhoods from frame to frame within the series of range images (also known as range image flow). An iterative method for computing the local translations is presented. In the global motion algorithm, a global motion model framework is utilized to compute a global motion estimation based on the polynomial expansion of the range images. We evaluate the algorithms on several real-world range image sequences with promising results.