【作者】 李长乐；

【导师】 赵杰；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2009， 博士

【摘要】 冷轧带钢表面缺陷检测技术是钢铁企业提高产品市场竞争力,改进生产工艺的关键技术之一。目前,传统的表面缺陷检测技术正逐渐被淘汰,取而代之的是基于机器视觉的表面缺陷检测技术,该技术已经成为带钢表面缺陷检测的主流技术。在20世纪末,欧美一些发达国家相继研究成功了基于机器视觉的表面缺陷检测系统。我国对这项技术的研究由于受到表面缺陷检测速度、图像处理速度等方面因素的制约,到目前为止尚无成功应用的系统,这在一定程度上影响了我国冷轧带钢产品的市场竞争力,不利于提高带钢产品的附加值。在机器视觉表面缺陷检测技术中图像处理算法繁多且复杂,而且当采用专用ASIC芯片或通用微处理器实现图像处理系统时,存在着灵活性和处理效率相互制约的问题。采用专用ASIC电路可以高速、可靠的实现图像处理,但是这种专用电路灵活性差,开发周期长,芯片设计比较复杂。通用微处理器(如计算机、DSP等)可以灵活的实现不同的图像处理算法,但是这种实现方法受到其处理器架构的制约,在实时性要求高的应用中,多采用并行处理器阵列来实现,使得系统庞大、价格昂贵、维护复杂。针对这些问题,本文研究了基于FPGA硬件平台的动态可重构技术,以及多IP核的图像并行处理技术,使得系统以较少的硬件资源实现比较复杂的功能,在提高系统执行速度的同时降低成本。分析了表面缺陷图像噪声来源及类型,在研究了基于偏微分方程图像去噪模型的基础上,选择了能够满足对比不变性和仿射不变性条件的AMSS(Affine Morphological Scale Space)方程对表面缺陷图像进行滤波处理,有效地去除缺陷图像中的噪声信息。采用有限差分法求解AMSS方程,并且分析了迭代步长和尺度参数对滤波效果的影响,改进了对称交叉熵的定义形式,提出了基于改进对称交叉熵的迭代停止准则,避免了依靠人为观察滤波结果,选择滤波尺度参数的问题,拓宽了基于AMSS方程图像滤波方法的应用领域,改善了表面缺陷图像预处理的效果。冷轧带钢表面缺陷由于受到带钢材质以及缺陷形成机理的影响,在缺陷与正常带钢之间普遍存在着过渡区域,同时该区域有助于识别带钢表面缺陷的类型,因而提出了基于过渡区的局部阈值图像分割方法。该方法充分利用了缺陷的过渡区信息,克服了局部阈值分割方法中子图像大小影响图像分割效果的问题,提高了缺陷分割的准确性和完整性。提取了缺陷的灰度特征、基于灰度共生矩阵的纹理特征和不变矩特征,采用核主成分分析的方法对缺陷同种特征进行抽取,降低了缺陷同种特征之间的相关性。基于信息融合理论,将缺陷特征组合,采用偏最小二乘法分析组合特征,降低组合特征中相同缺陷不同种类特征之间的相关性,使得组合特征能够更加准确、有效的描述缺陷。研究了支持向量机分类理论,分析了基于支持向量机的多类分类方法,将一类支持向量机多类分类的方法与不确定性理论相结合,设计了缺陷分类器并进行了分类实验,在有限样本的情况下,采用不同特征组合,使分类的准确率最高达到95%。研制了可重构图像并行处理系统,并且在系统中实现了表面缺陷图像的预处理和分割算法,提高了图像处理算法的运算速度,使得图像处理系统的实时处理速度可以达到39帧/秒。本文通过对表面缺陷图像检测技术理论和实验研究,研制了基于可重构技术和并行处理技术的表面缺陷图像处理系统,提出了表面缺陷图像处理及分类识别方法,在所研制的图像处理系统中实现了缺陷图像算法的软件硬件化,降低了图像处理系统的复杂度,提高了图像处理的执行效率,改善了表面缺陷检测系统的实时性。更多还原

【Abstract】 The cold rolling strip surface defect inspection technology is one of key technology, which could enhance the product market competitive strength of steel enterprise and improve the production engineering. The traditional surface defect inspection technology is being weeded out gradually, the surface defect inspection technology based on machine vision has replaced it and become the main trend of strip surface defect inspection technology. In 20 century’s ends, some developed countries in European and America have designed successfully the surface defect inspection systems based on machine vision. In our country the research on this technology is limited by surface defect inspection speed, image processing speed and so on, so far no successful system has been used on-line. This situation impacts the market competitive strength of our country’s cold rolling strip product partly and makes against for improving the strip product add-value.The image processing algorithm is various and complex in surface defect inspection technology based on machine vision. At the same time the image processing system has the interacting problem between flexibility and processing efficiency, which is realized by using specialized ASIC chips or general purpose MPU. The specialized ASIC chips could realize high-speed image processing reliably, but lacking flexibility and being unable to be modified once designed for certain algorithm function, also having long development cycle and complex chip-design. The general purpose MPU (such as computer, DSP and so on) could realize different image processing algorithms flexible, but limited by the processor’s architecture. The parallel processor array is used in the application with high processing speed demand, which makes hugeous system, expensive price and complex maintenance. Aiming at these problems, the dynamic reconfigurable technology based on FPGA hardware platform and the image parallel processing technology based on multi-IP core are researched in this paper, which make system realize more complex function with less hardware resource, improve the system execution speed and also reduce the cost.The source and types of surface defect image noise are analyzing. Based on researching on PDE image processing model, the AMSS equation meeting contrast invariance and affine invariance condition is selected for surface defect image smoothing processing, which removing the noise information of defect image effectively. The finite difference method is used to solve AMSS equation, also the iteration step length and dimension parameter’s impact on smoothing effect is analyzing. The definition format of symmetry cross entropy is improved. The iteration stop standard is brought forward based on improved symmetry cross entropy, avoiding the smoothing dimension parameter selection problem of depending on artificial observation of smoothing result and widening the application field of image smoothing method based on AMSS equation.The cold rolling strip surface defect is affected by strip material and defect forming mechanism, so transitional region exists between defect and normal strip widely. The transitional region information is helpful for the types identification of strip surface defect, so the local threshold value image segmentation method based on transitional region is brought forward. This method makes use of the transitional region information sufficiently, overcomes the problem that son image size could affect image segmentation effect in local threshold value method, enhances the accurateness and integrality of defect segmentation. Extracting the defects’gradation characteristics, veins characteristic and invariant moment characteristic based on gradation accrete matrix, the same defect characteristics are extracted using kernel PCA method, which reduces the interdependency among the same defect characteristics. Combinating defect characteristics based on information amalgamation theory and reducing the redundance among different defect characteristics in combination characteristic by using PLS to analyse combination characteristic, these make the combination characteristic describe the defect more precise and effective.The SVM classification theory is researched, the multi-class classification method based on SVM is analyzing. Combining a kind of SVM multi-class classification method with nondeterminacy theory, the defect classifier is designed and the classification experiment is done. On the situation of limited samples, the highest classification accurate ratio gets up to 95%, using different characteristics combination. The reconfigurable image parallel process system is developed, realizing surface defect image preprocessing and segmentation algorithm in the system, enhancing the mathematical operation speed of image processing algorithm. The real-time process speed of image processing system reaches to 39 frames per second.Through the theory and experiment research on surface defect inspection technology, the realization plan of reconfigurable parallel image processing system and the method of surface defect image processing and classification identification are proposed. The algorithm of defect image preprocessing and segmentation is realized in developed image processing system, reducing the complicacy of image processing system, enhancing the execution efficiency of image processing, improving the real-time ability of surface defect inspection system.更多还原