【作者】 刘伟嵬；

【导师】 颜云辉；

【作者基本信息】 东北大学 ， 机械设计及理论， 2009， 博士

【摘要】 冷轧带钢表面质量在线检测系统尽管已能够为钢铁企业提供现场监测,但自动识别系统中依然存在一些难点问题需要进一步深入分析和研究,如：检测系统的数据实时处理能力不足,检测系统的缺陷识别率不高等。本论文采集了大量冷轧带钢表面缺陷样本,围绕上述问题进行了深入的分析研究,提出了解决问题的方法,同时这些技术通过实验得到验证,具有一定的应用价值。本论文的主要研究内容及成果如下：(1)对冷轧带钢表面缺陷检测技术要求和应用特点进行分析,构建了基于客户端/服务器模式的分布式检测系统方案,有效地解决了多相机的图像处理与数据传输问题。提出了多线程处理和环形图像队列并用的图像实时采集方法,将图像采集线程和图像处理线程分开,使二者能够同时进行,有效地解决了图像数据堆积的问题,避免了因内存占用率过高而导致系统崩溃的情况。实验中,该设计使内存使用率基本稳定在528MB左右,CPU使用率基本稳定在14%左右,系统运行稳定,能够满足检测系统的实时性、准确性和可靠性的要求。(2)建立了评价带钢表面图像质量的基本原则,便于后续的图像处理,对于减小缺陷识别误差具有重要的意义。(3)提出了基于多元判别函数的疑似缺陷区域快速检测的算法,充分考虑了带钢表面缺陷的不确定性以及图像的纹理背景特征,应用多元判别分析实现带钢表面图像疑似缺陷区域的快速检测,使后续图像处理算法只对含有疑似缺陷图像进行处理,减少了系统的数据处理量。实验表明该算法的有效检出率可以达到99.5%,漏检率和误检率仅为0.5%和1.75%。(4)提出了基于邻域脉冲噪声评价(NINE)的脉冲噪声去除算法,在算法中设计了对图像噪声的判别准则,从而避免了图像细节的缺失,很好地保留图像原有的细节信息,解决了图像在滤除噪声后模糊不清的问题。与经典滤波方法相比,本文提出的NINE算法在脉冲噪声检测精度、客观失真测量以及视觉效果方面都有明显的提高。(5)提出了基于小波变换与非线性各向异性扩散技术相结合的带钢表面缺陷图像纹理背景信息滤除算法,该算法继承了小波变换和非线性各向异性扩散技术的优点,使滤波后的图像在边缘检测中基本消除了纹理背景对于缺陷检测的影响,能够获得清晰准确的缺陷边缘轮廓,有利于后续的缺陷自动分割等图像处理。同时,对经该算法滤波后的图像进行JPEG有损压缩,其平均压缩倍率是原始图像压缩倍率的2.3倍,为后续的图像处理、保存和传输节省了更多的存储空间。更多还原

【Abstract】 Though online surface inspection systems for cold rolled strips have been widely applied, some critical problems still need further research, such as insufficient real time processing, low defect recognization, etc. Plenty of defects samples of cold rolled strips were collected, and intensive study has been conducted with regard to the problems above. According, some new methods were proposed to settle these problems. All the technologies have been proved by experiments and would be valuable for industry application.Main works and achievements of the thesis are as follows:(1) A distributed detection system scheme based on Client/Server mode was used according to the analyses of technical requirements and application features for cold steel strip surface defect inspection. It has effectively solved the problem of multi-cameras image processing and image data transmission. A new image collecting method based on multithread processing and image ring queue was presented. This new method can separate image collecting thread from image processing thread and make them process simultaneously. Therefore the problem of image data accumulation was effectively solved and system collapse due to highly used memory was also avoided. In this case, the memory used stabilized at about 528MB and CPU utilization ratio stabilized at about 14%. The system operates stably and reliably.(2) General principles to evaluate rolled strips images were established which is quite useful to the following image processing. It is of great importance for reducing the error of defect recognition.(3) A new fast defect area detection algorithm using multivariate discriminant function was put forward, which takes full account of both defect uncertainty and image texture background. This algorithm makes it possible to only process the defective image in the following processing so that the amount of data processing will be reduced. In the experiment, the defect detection ratio can be up to 99.5%, and the misjudge ratio and omission ratio are separately 0.5% and 1.75%.(4) A neighborhood impulse noise evaluation was introduced. Image noise evaluation was added to this new algorithm, which was helpful to keep the image details and avoid blurring after removing the noise. Compared with classic filters, the new method performs effectively in noise detection accuracy, distortion measuring and vision effect.(5) A new filter based on wavelet transform and nonlinear anisotropic diffusion was proposed. With the advantage of both wavelet transform and nonlinear anisotropic diffusion, this new filter basically eliminated the effects of image texture background. Thus, clear and accurate outlines can be acquired, which is quite valuable to the following image auto-segmentation and other processing. At the same time, JPEG lossy compression was conducted to the images filtered, and the average compression rate is 2.3 times as that of the image without filtering, which saves more storage space for the image processing, saving and transmitting.更多还原