Probabilistic reasoning about ship images

One of the most important aspects of current expert systems technology is the ability to make causal inferences about the impact or new evidence. When the domain knowledge and problem knowledge are uncertain and incomplete, Bayesian reasoning has proven to be an effective way or forming such inferences ]3,4,8j. While several reasoning schemes have been developed. based on Ba.yes Rule, there has been very little work examining the comparative effectiveness of these schemes in a real application. This paper describes a knowledge based sys­ tem for ship classification Ill, originally developed using the PROSPECTOR updating method ]2j, that has been reimplemented to use the inference procedure developed by Pearl and Kim ]4,5]. We discuss our reasons for making this cha.nge, the implementation of the new inference engine, and the comparative performance or the two versions or the system. INTRODUCTION Classifying images is extremely difficult whenever the feature information available is incomplete or uncertain. Under such cir· cumstanees, identification of an object requires some kind or reuoning mechanism to help resolve ambiguous interpretations within the constraints of the available domain knowledge. The need lor a reasoning mechanism becomes even more acute ir the interpretation process is also constrained by limited resources. When there is not enough t.ime or memory for an exhaustive feature analysis, intelligent decisions must be made about how to use the resources available to maximum advantage. This means that the reasoning mechanism must be involved in the control or the information extraction activi· ties, as well as the interpretation of the results. Ship classification is an example of one practical application in which all or these problems a.rise. Classification of ships in an operational environment is a. difficult task regardless of what kind or images are used. This is not always obvious to those who are only rami­ liar with the dttailed vitws or a ship one finds in a reference book. Observers in the field rarely have the luxury or an abundance or clear details to work with. Images are most often obtained during a brief observa­ tion interval from a distance that makes high resolution difficult to achieve. The viewing angle is usually a matter or opportunity rather than choice, and the observer must make do with the prevailing visibility, weather, and lighting condition.s at sea. Another factor degrading image quality is 29 the fact that sensor platforms are often buffeted by turbulence in the air or the ocean. The qqality of images produced in this way is likely to be lower than that attainable using sophisticated enhancement techniques and powerful computing resources. These difficulties are of course exacerbated when the class ification must be done in real time. All or this is in addition to the complexity faced when distinguishing among hundreds or classes or vessels, some or which differ only in fine feature details. Having this task performed well is obvi­ ously important to the Navy, which has invested heavily in training personnel to analyze and interpret images under opera­ tional conditions. The human observer­ sensor operator must be highly trained and experienced. He/she must know which features are related to which ship classes, and make a judgement as to how well vari­ ous features are manifested in the image. Moreover, the observer must keep track of the implications or all these judgements both with respect to their uncertainty and consistency, and with respect to an eventual classification. A decision aid must also cope with these problems, but in a way that ack­ nowledges the meager computational resources available on most military plat­ forms an important constraint now and in the near future. The most useful kind of sys­ tem is one that can distinguish nmiltJr ship types. Most trained personnel can easily tell the difference between an aircraft carrier and a cruiser. It is much more difficult to make decisions about several types or cruisers whose images are similar. Real-time ship classification is a demanding application. It is a task requiring that complex inferences, based on incomplete and uncertain information, be made reliably under stringent computational constraints. In· devising a system that meets this chal­ lenge, two of the most important research issues are control and inference. Given that time constraints often preclude an exhaustive feature analysis, which features should be sought after in the time available! Given an uncertain and incomplete feature description, what kind or heuristic reasoning tools pro­ vide reliable and computationally inexpensive ship classifications! This paper describes a knowledge-based system for reasoning about ship images that successfully manages many ol these iasues. A prototype developed at the Navy Center lor Applied Research in AI (NCARAI) has convincingly demonstrated that a heuristic approach to this problem is effective and practical. Our current research effort builds on this work, and is developing a 2nd generation expert system to help solve this classification problem. REASONING ABOUT SHIP IMAGES The locus of this research is on how to use incomplete and uncertain feature infor­ mation to make plausible inferences about Naval Class. • Fipre 1 Examples or plan view reatures Reasoning about plausible classifications for a ship image requires knowledge about the features needed to describe various Naval Classes; and, knowledge about how the pres­ ence or absence or these features in an image implies one class versus another. Feature details might be observable from either a profile (or side) view, a plDn (or top down) view, or both. Figure 1 shows the kinds or features that are important in analyzing a plan view image. The primary items of • A Naval Cl188 is a &TOUp of shipe built to the same design and known by the lead ship's name.