面向带钢表面缺陷图像的特征提取算法研究

发布时间：2018-12-20 13:36

【摘要】：带钢作为钢铁产业的主要产品之一,广泛地应用于国民经济的生产和生活中。由于受到多方面因素的制约,带钢表面容易出现多种类型的缺陷。因此,及时、准确地检测出带钢的表面缺陷,对提高带钢产品的品质具有重要的意义。带钢在生产线上运行的速度较快,对缺陷检测系统的实时性要求较高。在实际的生产环境中,采集到的带钢表面图像的绝大部分都是无缺陷图像,因此可以在特征提取操作之前采用快速预判技术过滤缺陷图像。快速预判包括二值化和缺陷判断两个阶段,二值化处理能够更清晰地呈现缺陷部分,然后缺陷判断统计图像中的灰度分布特征即可确定图像中是否存在缺陷。在图像分析中,纹理是一种重要的特征。考虑到大部分带钢缺陷图像中都存在丰富的纹理信息,利用灰度共生矩阵进行特征提取,生成方向、生成步长和灰度级这三个构造因子被用于构造特征向量。同传统方法将特征参数在不同方向上的平均值作为最终特征向量的元素不同,针对带钢图像的特征提取将每个方向所得到的多个特征参数的平均值作为最终特征向量的元素,实验证明该方法显著提高了带钢缺陷的分类准确率。方向梯度直方图(Histogram of Oriented Gradient,HOG)是另一种特征提取的方法。最初HOG用于检测静态图像中的行人,后由于其极强的图像特征描述能力被扩展到各个领域。针对带钢缺陷分类问题,对影响HOG特征的参数进行了优化,并改进了HOG特征的梯度计算模板,实验证明高维的模板更加有利于带钢缺陷的分类。
[Abstract]:As one of the main products of steel industry, strip steel is widely used in the production and life of national economy. Due to the restriction of many factors, many kinds of defects are easy to appear on the strip surface. Therefore, timely and accurate detection of strip surface defects is of great significance to improve the quality of strip products. The speed of strip steel running on the production line is fast, and the real-time requirement of defect detection system is high. In the actual production environment, most of the images collected from the strip surface are non-defect images, so we can filter the defect images before feature extraction operation. The fast pre-judgment includes two stages: binarization and defect judgment. The binary processing can show the defect part more clearly, and then the gray-scale distribution characteristics in the image can be determined to determine whether there are defects in the image. Texture is an important feature in image analysis. Considering that there is abundant texture information in most strip defect images three constructional factors such as step size and gray level are used to construct feature vectors by using gray level co-occurrence matrix to extract features generate direction and generate step size and gray level. Different from the traditional method, the average value of the feature parameters in different directions is taken as the element of the final feature vector, and the average value of multiple feature parameters obtained from each direction is taken as the element of the final feature vector for the feature extraction of the strip image. Experiments show that this method can improve the classification accuracy of strip defects. Directional gradient histogram (Histogram of Oriented Gradient,HOG) is another feature extraction method. HOG was initially used to detect pedestrians in static images, but it was extended to various fields due to its strong feature description ability. Aiming at the problem of strip defect classification, the parameters affecting HOG feature are optimized, and the gradient calculation template of HOG feature is improved. The experiments show that the high dimensional template is more favorable to the classification of strip defects.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】