钢板表面缺陷图像检测与分类技术研究

发布时间：2018-05-16 22:30

本文选题：钢板缺陷检测 + 不均匀光照矫正　；参考：《大连海事大学》2017年硕士论文

【摘要】：钢材行业作为国民经济的基础产业,在经济建设、社会发展等多方面都发挥着重要作用,但钢板表面缺陷严重降低钢材质量,钢板表面缺陷检测是保证钢材质量的关键因素之一。本文对钢板表面缺陷检测与分类算法进行了较为深入的研究,具体成果如下:(1)针对钢板表面图像数据量大、计算效率低的问题,本文构造了基于梯度的钢板表面图像感兴趣区域检测算法。通过钢板表面图像的局部梯度统计值判断有无缺陷,该算法可以减小系统后端图像处理压力,提升整体效率。(2)针对部分缺陷图像受不均匀光照影响且对比度低的问题,本文构造了基于Retinex算法与导向滤波器的不均匀光照矫正和增强算法。利用导向滤波器估计光照分量,根据Retinex算法计算反射分量,并对反射分量进行增强,提高对比度,恢复原始灰度信息。(3)针对缺陷分割准确度问题,本文构造了基于Center-Surround Difference的缺陷图像分割算法。利用带权重的组合DoG(Difference of Gaussian)滤波器对预处理后缺陷图像进行滤波,根据Center-Surround Difference提取图像Local特征和Global特征,并将两者线性融合,对融合后图像进行背景抑制和前景恢复,利用自适应阈值实现缺陷区域提取。(4)根据缺陷图像分割前后所含的不同信息,提取36维特征值作为缺陷分类依据。针对缺陷分类精度问题,从惯性权值和种群多样性两个方面对粒子群(Particle Swarm Optimization,PSO)算法进行改进,利用改进算法对前馈(Back Propagation,BP)神经网络进行优化,最终实现缺陷分类。实验结果表明,本文设计的算法能很好的实现缺陷的检测和分类识别,实验主观效果和客观效果保持一致。
[Abstract]:As the basic industry of the national economy, the steel industry plays an important role in economic construction, social development and other aspects, but the steel plate surface defects seriously reduce the quality of steel, Surface defect detection of steel plate is one of the key factors to ensure steel quality. In this paper, the surface defect detection and classification algorithm of steel plate is studied in depth. The concrete results are as follows: (1) aiming at the problem of large amount of image data and low computational efficiency of steel plate surface, In this paper, a gradient based region of interest detection algorithm for steel plate surface images is proposed. Based on the local gradient statistical value of the steel plate surface image, the algorithm can reduce the image processing pressure and improve the overall efficiency of image processing, aiming at the problem that some defective images are affected by uneven illumination and the contrast is low. In this paper, a nonuniform illumination correction and enhancement algorithm based on Retinex algorithm and guide filter is constructed. The illumination component is estimated by the guide filter, the reflection component is calculated according to the Retinex algorithm, and the reflection component is enhanced, the contrast is improved, the original gray level information is restored. A defect image segmentation algorithm based on Center-Surround Difference is proposed in this paper. A combined DoG(Difference of Gaussian filter with weights is used to filter the pre-processed defective image. The Local and Global features of the image are extracted according to Center-Surround Difference, and the two features are linearly fused to suppress the background and restore the foreground of the fused image. Based on the different information before and after the defect image segmentation, 36 dimensional eigenvalues are extracted as the basis of defect classification. Aiming at the accuracy of defect classification, the particle swarm optimization (PSOs) algorithm is improved from the aspects of inertia weight and population diversity. The improved algorithm is used to optimize the feedforward back Propagation (BP) neural network, and finally the defect classification is realized. The experimental results show that the algorithm designed in this paper can achieve the defect detection and classification, and the subjective and objective effects of the experiment are consistent.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】