A recent trend in Machine Vision has been the use of increasingly accurate physical models for the development of image understanding algorithms. These models typically describe the relationship between the image of a surface and the illumination environment, the shape and optical properties of the surface, and the characteristics of the image sensor. The physics-based approach carries several important advantages. In many instances, the study of physical models suggests algorithms that exploit image structure that would otherwise be ignored or misinterpreted. The incorporation of explicit physical models in algorithms allows accurate characterization of the class of images that can be processed with a given degree of accuracy. This characterization is important in guiding potential users in the selection of algorithms for applications. Since physical models relate image observables to illumination and sensor properties, high performance algorithms can be derived that rely on specialized sensors or carefully controlled illumination conditions.

During the 1970s, only a small minority of machine vision researchers led by Berthold Horn and his colleagues at MIT embraced the physics-based approach. With time, however, others recognized the importance of physical models and a sizable community of physics-based vision researchers had formed by the late 1980s. This group's necessarily interdisciplinary approach to machine vision led to the publication of results in a wide range of conferences and journals. In the summer of 1992, much of this work was reviewed during a symposium on physics-based vision and a three volume collection of papers covering the major subareas of radiometry, color, and shape recovery was published (Physics-based Vision: Principles and Practice, Jones and Bartlett, 1992). These volumes summarized the success of physics-based vision in areas such as reflection and sensor modeling, color segmentation, color constancy, shape from shading, and photometric stereo.

In this feature, we present fourteen papers that describe recent progress in physics-based machine vision. This sampling of papers is reasonably indicative of the current breadth of research in the field. The first five papers contribute to sensor and reflection modeling and include two papers that propose new classes of sensors. The next five papers present recent results in multispectral processing with papers considering both visible and infrared regions of the spectrum. The final group of papers describes advances in systems that recover surface shape information using images obtained under different illumination configurations. Nearly all of the papers in the feature present experimental examples to illustrate the effectiveness of models and the performance of algorithms.

Back to the ICVL homepage