Occluded object recognition based on the theory of evidence.

摘要

In this project, we have designed a remedial method whenever the current computer vision system for I. C. manufacturing industry encounters a failure under non-uniform lighting. To achieve this goal, we use edges or line segments instead of pixel intensities. These high-level features obviously are more robust to non-uniform lighting. However, it is not always possible to have perfect edge detection or line segmentation. Some of them will be lost and some will be distorted after the feature extraction process. This brings us to face the occluded object problem. In fact, the question becomes: What is the adequate amount of detected features that is good enough for us to consider the wanted object being present? In this research, we have solved some critical problems in order to answer this fundamental question.;Our approach will directly calculate the probability of the wanted object being present with the amount of detected features. If this probability is high, we will consider that the wanted object is present. To deduce this probability, we have to perform several processing steps. First, we need form a finite set with all the possible objects. We then quantify the evidences (obtained from detected features) in term of probability for supporting the presence of the objects in the finite set using the Bose-Einstein model. Afterwards, we combine all these evidences using the Dempster and Shafter theory to deduce the probability of the wanted object being present.;As the evidences derived from the objects in the finite set can be dependent, the Dempster rule for combining independent evidences needs modifications. However, when the evidences become dependent, it is very complicated to calculate the combination result exactly. In order to simplify the algebraic manipulations, we suggest to use the minimum and maximum operations instead of multiplication to calculate the worst case for the presence of the wanted object. As a result, we have derived a generalized equation for a group of objects in the finite set.;For industrial applications, we choose the recognition of IC dies. The matching score, in term of probability, will be lower than 0.5 when the amount of detected features is less than that of common features with the wanted object. Hence, our approach has a high discrimination power. In a practical environment, we have to consider the processing speed. Our algorithm is much complicated than the normalized cross correlation. When it is implemented with software, it requires about 10 minutes for a 256 x 256 image. Hence, our method is too slow when compared to the current normalized cross correlation which can attain real-time speed with special hardware. However, we will only apply our algorithm when the current system fails, and it is still much faster than a skillful worker in such situation.