Image Understanding in Unstructured Environment

The following sections are included:

• PREFACE

• CONTENTS

The following sections are included:

• INTRODUCTION

• OVERVIEW OF THE DESIGN
  • Representation
  • Control

• THE CKS DATABASE
  • Data Storage and Retrieval
  • Data Tokens

• ORGANIZATION OF SPATIAL INFORMATION
  • Spatial Directory
  • Retrieval
  • Implementation

• ORGANIZATION OF SEMANTIC INFORMATION
  • The Vocabulary
  • Attributes
  • Semantic Relations
  • The Semantic Directory

• LOGICAL INTERPRETATION OF THE DATABASE
  • Semantics
  • Insertions
  • Queries
  • Discussion

• PROCESS CONTROL
  • Daemons
  • Meta-Level Control

• EXTENDED EXAMPLE

• SUMMARY

• References

A prototype system for classifying complex ship images has convincingly demonstrated that Bayesian reasoning is a valuable tool for making plausible inferences about classificatory hypotheses given impoverished feature data¹. It remains to be shown that such methods are also useful in handling the large scale, resource-constrained classification problems that are of interest to the Navy. Classifying objects using sensor data inan operational environment is a demanding task. Regardless of the kind of sensor information available – visual,infrared,radar,or sonar – this is a task in which complex inferences must be made reliably under stringent computational constraints, and based on incomplete and uncertain evidence. This paper describes research efforts focused ondevising a robust and accurate classification problem solver that meets this challenge.

Extraction of 3-D motion and structure from a monocular image sequence is considered. The approach is essentially a regional technique and is different from the popular methods that are based on point correspondence. The technique requires that each image in the sequence be first segmented into regions, with each region corresponding to certain visible planar face in the object. The dynamic nature, i.e., the evolution of these shapes are used to extract the 3-D motion parameters, while certain static (intraframe features) are utilized to estimate the initial 3-D structure of the object. For orthographic images the shape change is viewed as a linear transformation, while it is nonlinear for the central projection.

A set of methods of computing these shape transformation coefficients are presented. First, a class of iterative methods based on operator formulations are considered. The second approach is based on the moment invariants. A combination of the above two, with reasonably chosen heuristics is also adopted to accelerate the search. In addition, certain mutual constraints among these coefficients are derived, which play the role similar to that of Kanade’s results on shape recovery. We present techniques of extracting 3-D orientation and position from three views of one face (region) or from two views of two faces. To emphasize the distinction from the point correspondence approach, we also examine in detail similar regional or contour based approaches that are adopted by others.

Human perception has long inspired computer vision and image understanding research. In this paper we present a spherical vision model which is inspired by human perception and which is useful in the study of computer vision and image understanding. More precisely, we discuss spherical geometry, creating and displaying spherical images, and applications of spherical methods to problems in surface recovery, motion analysis, and optical flow.

A spatial reasoning and decision support system for remote sensing is being developed. An image processing and understanding module is employed to process input sensory dataand to extract features so that its classification results can be used in other spatial information processing modules, such as reasoning and decision support. In this paper, three Bayesian decision theoretic approaches to contextual classification are developed in the image processing and understanding module which make use of spatial information. These approaches are then applied to understanding of remote sensing imagery.