Machine Vision Recognition Technology Study Of Structure Crack

Machine vision is the interdisciplinary field of optical engineering, digital image processing, pattern recognition, robotics and other areas, with advantages of non-contaction, real-time, high visualization. Among which intelligent algorithms are applied to replace human to conduct automation detection work without influence of environment, monitoring the targets and making timely decisions which has been widely used in many fields. Aiming to improve the crack recognition accuracy, theory analysis and experiment validation in machine vision and digital image processing are both conducted, and the principle of Image pre-processing, Hough transform, and Support Vector Machin e are studied in-depth. Connectivity feature of crack line is verified experimentally and optimize the entire recognition algorithm further. Main content is as following:(1)Crack image pre-processing. Basic methods and principle are described. Adopt the proper filtering algorithm based on real collected crack images, thus lay the foundation for subsequent crack feature extraction and classification of sample training.(2)Feature extraction algorithm by improved Hough transform. Main principle of Gradient response, Hough transform on line and circle detection are described. Taking actual crack characteristics into account, Hough transform algorithm is re-designed to recognize crack. Apply the principle of Hough transform in binary image into gray image, extend accumulator in algorithm from two value 0 and 1 to any numeric value between 0 and 1, so crack features can be extracted more precisely.(3)Classification training and regional connection of crack. Support Vector Machine is used to class and training ma ssive crack samples, and further result forecast, and label crack area and connect crack line. Principle of classification and Kernel function based on machine learning is introduced. Train lots of different samples to obtain SVM train model. SVM predict w hether a cell is crack according to feature matrix. Then crack area is labeled by rectangle. On this base, remove isolated labels, connect the others with line using the algorithm of One Component at a time in 4 neighborhood.(4)Experiment verification. On e is crack area labeling and line connection experiment with different size(200*200 pixel and random size) and different numbers of image. The other is comparative analysis of 4 different recognition algorithm based on different features. Use Matlab and O pen CV software, verify the algorithm accuracy with Visual Studio program. Results table and ROC curve are used to record the comparative results. The results show that, the proposed algorithm has high accuracy, which can meet requirements of crack recognit ion with machine vision.