Research On Recognition Based On Quasi-circular Objects

In real life, quasi-circular objects generally exist everywhere. The recognizing of this type of objects is called quasi-circular recognition which holds an important position in computer recognition. Because recognizing quasi-circular objects by computer technology has some advantages: easing work intension and enhancing the efficiency and accuracy of work, many specialists and scholars have done research in this field and made noteworthy progress. But at present there is no a uniform method of recognizing the quasi-circular objects because of their characteristics, furthermore, the results of existing methods are not satisfactory in real time and precision performances.This thesis tries to recognize the quasi-circular objects by using the theory of digital imaging and pattern recognition. The main research includes the following areas: image pre-processing, image segmentation and image recognition. A new recognition method based on watershed segmentation is brought forward. This method firstly segments the objects by watershed transform, and then combines cluster analysis and fuzzy recognition method to accomplish the detecting of objects. Owning to the simple arithmetic and logic operation, this method has real time property, and the total CPU time is less than 1 second for one 640*480 pixels image. The grayscale-weighted threshold algorithm is used to binarize the image. This algorithm can choose different weights according to the brightness and the degree of cauterization of each image, which enhances the quality of the binary image and effectively reduces the frequency of excessive segmentation of watershed transform. The recognition method based on scanning is improved in this thesis, and is combined with the watershed transform segmentation algorithm to make a new method called watershed-scanning recognition method, which enhances the precision and efficiency of recognition.In the experiment part, steel bar and log images are recognized by the two methods put forward in this thesis, and then algorithm comparison and error statistic are present. From the experiment results, the relative error of the two methods is less than 3%.