Research On Counting Method Of Bundled Bars Based On Machine Learning

This thesis presented a method of bars counting based on machine learning aming at the counting problem of bundled bars.We compared and analyzed these results which studied and obtained by these exiting counting methods and the method proposed in this thesis.The main work includes three aspects as follows.Firstly,a modified template matching method with muti-threshold was presented aiming at the gray difference of bottom surface in the image.Gray image was divided into multiple binary images by threshold.Match multiple binary images with standard binary image,which was called template,and get the number of bars according to the matching results.The experimental results showed that The multi-threshold template matching method proposed in this thesis is more suitable for the situation where the gray of bottom surface is different than that of single threshold and adaptive threshold template matching method.Secondly,a SVM method was proposed to solve the problem of bars counting while traditional counting methods have some limitations.HOG feature of each detecting windows was constructed and used in the SVM classifier with Gaussian kernel.The center position of each bar bottom surface was labeled manually and the training sample set was constructed according to these standard mark points.And some results of template matching were selected and used to make up the sample set.Then the SVM classifier was trained which was used to determine whether the detecting window includes a bar or not.The counting result was obtained by counting the number of connected regions in the image of marked centers.The experimental results showed that the counting accuracy of SVM classifier with gaussian kernel was improved compared with the counting results of template matching methods.Thirdly,sufficient and comprehensive samples make a great influence on the performance of the SVM classifier.An iteratively training SVM method was proposed to strengthen the performance of trained SVM classifiers aiming at the problem that the samples involved in the training are not typical.The misclassified samples were used to construct the next training sample set.And counting on the training image library and test image library.The experimental results showed that this strategy improved the performance of SVM classifier effectively in training image library and test image library.The accuracy of this method was greatly improved compared with the conventional methods,and it had better robustness.In conclusion,this thesis had conducted a beneficial study on the counting method of bundled bars,and the proposed method had certain reference value for improving the accuracy of counting problem.