Louis A. Tamburino

Automatic generation of morphological sequences

Morphological sequences (algorithms or programs) are generated using an evolutionary approach. A population of morphological sequences is manipulated and expanded in discrete steps. At each time-step two tasks are initiated--program discovery and program construction. The discovery phase searches for short morphological sequences which extract novel features. Program composition utilizes these sequences, which are partial solutions, to form increasingly effective sequences. The composition phase selects pairs of sequences and combines them into extended sequences which capture spatial relationships. The enhanced population serves as the basis for another phase of discovery and composition. Several demonstrations illustrate the systems ability to synthesize and integrate feature extraction routines.

Adaptive search for morphological feature detectors

A closed-loop hybrid learning system that facilitates the automatic design of a multi-class pattern recognition system is described. The design process has three phases: feature detector generation feature set selection and classification. In the first phase a large population of feature detectors based on morphological erosion and hit-or-miss operators is generated randomly. From this population an optimized subset of features is selected using a novel application of genetic algorithms. The selected features are then used to initialize a generalized Hamming neural network that performs image classification. This network provides the means for self-organizing the set of training patterns into additional subclasses this in turn dynamically alters the number of detectors and the size of the neural network. The design process uses system errors to gradually refine the set of feature vectors used in the classification subsystem. We describe an experiment in which the hybrid learning paradigm successfully generates a machine that distinguishes ten classes of handprinted numerical characters.

Multi-resolution feature extraction from Gabor filtered images

In this paper, we describe a hybrid learning system which combines a genetic algorithm with a neural network to classify grayscale images. The system operates on multi-resolution images which are formed by applying Gabor filters to a set of input images. The genetic algorithm evolves morphological probes that sample the multi-resolution images, and the perceptron algorithm then evaluates the extracted features. We demonstrate the use of this system by discriminating images of model tanks from other military vehicles. A multiplicity of accurate solutions, consisting of sparse morphological probes, are generated.<<ETX>>

Approaches to synthesizing image processing programs

Machine learning techniques are examined as a means of automatically generating image processing programs. Nonstructured techniques such as discovery systems and evolutionary processes are studied because they facilitate the exploration of enormous search spaces without a detailed knowledge base. The success of these methods depends on the algorithm representation and the effectiveness of performance evaluation. Mathematical morphology provides an algebraic representation which is powerful and challenging to program. The qualitative aspects of effective performance measures are also discussed.<<ETX>>

Evolving robust gender classification features for CAESAR data

In this paper we explore the robustness of histogram features extracted from 3D point clouds of human subjects for gender classification. Experiments are conducted using point clouds drawn from the Civilian American and European Surface Anthropometry Resource Project (CAESAR anthropometric database provided by the Air Force Research Laboratory (AFRL) Human Effectiveness Directorate and SAE International). This database contains approximately 4400 high resolution LIDAR whole body scans of carefully posed human subjects. Features are extracted from each point cloud by embedding the cloud in series of cylindrical shapes and computing a point count for each cylinder that characterizes a region of the subject. These measurements define rotationally invariant histogram features that are processed by a classifier to label the gender of each subject. Preliminary results using cylinder sizes defined by human experts demonstrate that gender can be predicted with 98% accuracy for the type of high density point cloud found in the CAESAR database. In our previous has shown that when point cloud densities are reduced to levels that might be obtained using stand-off sensors; gender classification accuracy degrades. In this paper we show the results of how the classification accuracy degrades as a function of center of mass displacements.

Evolving point-cloud features for gender classification

In this paper we explore the use of histogram features extracted from 3D point clouds of human subjects for gender classification. Experiments are conducted using point clouds drawn from the CAESAR anthropometric database provided by the Air Force Research Laboratory (AFRL) Human Effectiveness Directorate and SAE International. This database contains approximately 4400 high resolution LIDAR whole body scans of carefully posed human subjects. Features are extracted from each point cloud by embedding the cloud in series of cylindrical shapes and computing a point count for each cylinder that characterizes a region of the subject. These measurements define rotationally invariant histogram features that are processed by a classifier to label the gender of each subject. Preliminary results using cylinder sizes defined by human experts demonstrate that gender can be predicted with 98% accuracy for the type of high density point cloud found in the CAESAR database. When point cloud densities are reduced to levels that might be obtained using stand-off sensors; gender classification accuracy degrades. We introduce an evolutionary algorithm to optimize the number and size of the cylinders used to define histogram features. The objective of this optimization process is to identify a set of cylindrical features that reduces the error rate when predicting gender from low density point clouds. A wrapper approach is used to interleave feature selection with classifier evaluation to train the evolutionary algorithm. Results of classification accuracy achieved using the evolved features are compared to the baseline feature set defined by human experts.

Efficient software techniques for morphological image processing with desktop computers

The basis of a system for processing binary images with the operations of mathematical morphology is described. This system exploits the properties of mathematical morphology to minimize computing time and storage requirements. Images are stored in data structures which are memory-efficient and allow several images to be processed simultaneously. Techniques are also presented for efficiently storing globally sized structuring elements. These ternary images are stored in data structures which utilize an adaptive window to provide storage for a 2M X 2N specification space in an optimal M X N data structure. This representation provides efficient storage, retrieval, and comparison of generalized structuring elements.

Electro-optical seasonal weather and gender data collection

This paper describes the process used to collect the Seasonal Weather And Gender (SWAG) dataset; an electro-optical dataset of human subjects that can be used to develop advanced gender classification algorithms. Several novel features characterize this ongoing effort (1) the human subjects self-label their gender by performing a specific action during the data collection and (2) the data collection will span months and even years resulting in a dataset containing realistic levels and types of clothing corresponding to the various seasons and weather conditions. It is envisioned that this type of data will support the development and evaluation of more robust gender classification systems that are capable of accurate gender recognition under extended operating conditions.

The effects of clothing on gender classification using LIDAR data

In this paper we describe preliminary efforts to extend previous gender classification experiments using feature histograms extracted from 3D point clouds of human subjects. The previous experiments used point clouds drawn from the Civilian American and European Surface Anthropometry Project (CAESAR anthropometric database provided by the Air Force Research Laboratory (AFRL) Human Effectiveness Directorate and SAE International). This database contains approximately 4,400 high-resolution LIDAR whole body scans of carefully posed human subjects. Features are extracted from each point cloud by embedding the cloud in series of cylindrical shapes and computing a point count for each cylinder that characterizes a region of the subject. These measurements define rotationally invariant histogram features that are processed by a classifier to label the gender of each subject. The recognition results with the tightly control CAESAR database reached levels of over 90% accuracy. A smaller secondary point cloud data set was generated at Wright State University to allow experimentation on clothed subjects that was not possible with the CAESAR data. We present the preliminary results for the transition of classification software using different combinations of training and tests sets taken from both the CAESAR and clothed subject data sets. As expected, the accuracy achieved with clothed subjects fell short of the earlier experiments using only the CAESAR data. Nevertheless, the new results provide new insights for more robust classification algorithms.

Exploring point-cloud features from partial body views for gender classification

In this paper we extend a previous exploration of histogram features extracted from 3D point cloud images of human subjects for gender discrimination. Feature extraction used a collection of concentric cylinders to define volumes for counting 3D points. The histogram features are characterized by a rotational axis and a selected set of volumes derived from the concentric cylinders. The point cloud images are drawn from the CAESAR anthropometric database provided by the Air Force Research Laboratory (AFRL) Human Effectiveness Directorate and SAE International. This database contains approximately 4400 high resolution LIDAR whole body scans of carefully posed human subjects. Success from our previous investigation was based on extracting features from full body coverage which required integration of multiple camera images. With the full body coverage, the central vertical body axis and orientation are readily obtainable; however, this is not the case with a one camera view providing less than one half body coverage. Assuming that the subjects are upright, we need to determine or estimate the position of the vertical axis and the orientation of the body about this axis relative to the camera. In past experiments the vertical axis was located through the center of mass of torso points projected on the ground plane and the body orientation derived using principle component analysis. In a natural extension of our previous work to partial body views, the absence of rotational invariance about the cylindrical axis greatly increases the difficulty for gender classification. Even the problem of estimating the axis is no longer simple. We describe some simple feasibility experiments that use partial image histograms. Here, the cylindrical axis is assumed to be known. We also discuss experiments with full body images that explore the sensitivity of classification accuracy relative to displacements of the cylindrical axis. Our initial results provide the basis for further investigation of more complex partial body viewing problems and new methods for estimating the two position coordinates for the axis location and the unknown body orientation angle.