Bundled Round Bars Counting Based on Iteratively Trained SVM

摘要

Bundled bars counting is difficult in the cases of overlap or varied illumination. An iteratively trained SVM method is proposed to count bundled round bars from a bottom side image. Using Hough transformation, the sizes of bars are extracted and normalized. A SVM classifier using HOG features of the image are applied to determine the center points of bars. These center points generate central regions corresponding to bars. By counting the number of connected regions with great area in the image, the number of bars is obtained. In SVM training process, sample selection affects the classifier significantly. From an iteratively selection process, typical samples are selected and used for training the SVM classifier. The experimental results showed this strategy improved the performance of SVM classifier effectively, and the method works well in overlapped or varied illumination situation.