Since the start of the information age, pioneers in computing and mathematics have visualized machines exhibiting human-like intelligence. Eventually, it was believed, machines would not only replace humans performing all manner of dangerous or repetitive tasks, but aid them in solving the most difficult problems of the day. By all accounts, such machines have been slow to materialize.
Human beings have a propensity for the visual senses. Of all the fields of machine intelligence, it is machine vision which has amassed the greatest number of solutions, stretching as far back as the late 1950's. Machines which can quickly recognize or understand digital images are in great demand for numerous industrial and military applications:
For a machine vision system to be useful, it must intelligently emit some kind of decision about the environment it is operating in, based on images aquired from cameras, infrared sensors, or other physical transducers. To arrive at a such a decision in minimal time, it is desirable to reduce the amount of data which the system is required to process. Through a process of feature extraction the input imagery can be reduced to a few components which are essential to a particular problem.
The evolution of machine vision has created a great many diverse methods for extracting relevant features from images. This paper focuses on a particular feature extraction method which employs moment invariance to achieve good results when the observed physical environment is sufficiently constrained. With proper lighting, acquired 2D intensity images can be thresholded to yield contiguous swatches of pixels called blobs, and each blob can then be treated as a unique region of interest on which the moment processing will be performed. Initial work in this area has been traced back to a paper written in 1962 by Ming-Kuei Hu, and to an implementation of Hu's derived invariants for aircraft recognition in the 1977 paper by Dudani et. al.
In this paper, a simple machine system which can detect a pair of scissors, a knife, and a meat cleaver using Hu's system of invariant equations is presented. Section 2 describes the general theory of moments and moment invariant functions as derived by Hu. Section 3 presents the simple machine vision system and relevant results. The paper is finished with a discussion on utilizing moment invariants in a modern machine vision context.