Modern Computer Vision History

6

History

Modern computer vision researches can be traced back to the early 1960s. The first applications to achieve were pattern recognition systems, which we aimed to recognize characters in office automation related tasks. The urge to match 2-dimensional features extracted from images with the 3-dimensional representation of objects first occurred in 1960s in early works by Roberts L. G. Roberts, Machine Perception of Three Dimensional Solids, Ph.D. thesis, Massachusetts Institute of Technology, 1963.

. Early investigations on vision systems can be also related to the Hitachi labs in Japan, where was originated the term machine vision. With the aim to replace human workers early investigation in 1964 included the automation of the wire-bonding

Nevertheless, it is proven that it is highly difficult to create a computer system with this ability under the complexity of this task. For example, when for the existing computer vision systems it is possible to detect flaws in products that would go unnoticed by most human observers, it is difficult to build a system that would posses an ability of three year old child to identify his own toys from his friend. This is because we can build a machine to look for certain features, but we don't know what features will be useful for identifying everyday objects. Other condition is that environment, in which the system should work, usually is controlled – like for example the size of input image, the light will be always the same, but meanwhile in a real world such circumstances appear quite rarely and are hard to predict. Objects can appear in different orientations, under different illuminations and different sizes. For example, let's imagine a clear shadow of a chair on the wall. Naturally, we would not make an attempt to sit on it; meanwhile the system will still recognize it as a chair. Furthermore, due to our ability of contextual learning, under certain circumstances even if a part of object is hidden we can still determine what we see. Under these conditions the general problems of object recognition are easy to

When the typical object recognition system is visualized in a picture (), the main problems occurring can be summed up to four:

Detection: is a particular item present in the stimulus?

Localization: detection plus accurate location of item

Recognition: localization of all the items present in the stimulus

Understanding: recognition plus role of the stimulus in the context of the scene

Outline

Regarding to main issues stated before, there are many methods for implementing object recognition.

Appearance – based methods

Appearance – based methods use images as an example of objects to perform recognition; the conditions could be changes in lighting or color, in viewing direction or size/shape.

Edge matching using Canny edge detection to find