连铸热坯表面缺陷异源CCD信息融合三维形貌精确识别方法研究

Nowadays continuous casting (CC) technology has become main solidification process for liquid steel and the supply side for steel rolling material, therefore, in harsh environment on-line monitoring for CC products surface quality is of great significance, which means by precise surface quality inspection smooth running of CC process can be guaranteed, meanwhile it has significance to energy saving, hot delivery , direct rolling rate and the final product quality. In this project through studying the allogenetic CCD laser scanning synchronization imaging and optimal fusion algorithm, we combine line scan CCD conventional imaging and area scan CCD 3D imaging strategy, it aim is to study and establish the description method for uncertain region of interest (ROI) by employing upper and lower approximation set set and boundary region of rough set theory. Based on this, through establishing fuzzy connectedness among pixels in the above three regions to construct a definitive fuzzy classification model for delineating the boundary of ROI. Thereinto, in order to improve the positioning accuracy and delineation precision of 3D shape for a specific ROI, first, by experimental study the imaging features of CC product surface oxidation scales,vibration marks,defects and normal area need to be analyzed by optical computing,further, to study the formation mechanism of noise in recognition process and the suppression method for pseudo defect interference. Second, the fast image registration and fusion algorithm need to be determined by analyzing and computing the correlation and mutual information feature between allogenetic CCD images. Finally,to study Muti-ROI segmentation algorithm for common defects and training method for the delineation parameters of fuzzy connectedness , which aims to acquire the accurate defect 2D shape,moreover,based on allogenetic CCD image fusion information to quantize the depth feature of each detected ROI. The contributions of this project can provide theoretical basis and technical support for developing new vision-based defect inspection system and improve recognition rate of existing visual-based inspection system.

高噪声下连铸热坯表面质量在线监控，对保障连铸顺行、节省能源及提高热送直轧率具有重要意义。项目以铸坯表面缺陷在线检测为背景，以异源CCD线扫描、激光调制面扫描及信息融合为研究内容。首先，实验测定铸坯表面氧化膜、常见缺陷及背景成像特征，研究CCD成像噪声形成机理和伪缺陷干扰抑制方法；其次，研究建立基于粗糙集的铸坯表面感兴趣区域(ROI)上、下近似及不确定边界描述方法。以像素相对模糊连接度为判别准则，构建区域归属竞争机制及模糊分类模型，对ROI二维形态进行精确分割；其三，通过分析异源CCD互信息特征，建立图像配准与底层数据融合算法，精确重构不完整缺陷形态，提取缺陷区域深度信息，提高缺陷定位和三维形态量化检测精度；最后研究建立相对模糊连接度铸坯表面多缺陷类型分割参数自适应学习和同步分割方法，提高热态下铸坯表面缺陷识别率。本项目研究成果为高噪声下开发新型视觉探伤装置提供理论和技术支撑。

本项目以连铸坯表面缺陷可靠检测为研究背景。连铸作为钢业凝固成型的主要技术工艺，已在钢铁冶金领域得到广泛推广，连铸无缺陷坯生产技术也成为研究热点。近些年，随着国家制造2025发展钢业的颁布，制造业正在经历的从传统制造向智能制造的转型升级。在钢业冶金行业连铸直轧技术和热送热装(HC-DR)生产工艺正在经历快速发展阶段，其中，无缺陷坯生产技术该些技术推广实施的必要条件。但由于连铸生产过程存在多种非稳稳态因素，很难从连铸生产工艺本身确保无缺陷坯生产。因此，无损非接触在线检测技术逐渐成为了HC-DR生产工艺的中间支撑环节，.本项目以机器视觉智能检测技术为主要研究内容，开展高噪声下连铸表面缺陷可靠检测，为连铸坯HC-DR提供技术支撑，研究内容涉及连铸坯缺陷检测双目视觉三维激光扫描成像方法，基于粗糙集和模糊连接度的表面缺陷精确识别方法以及基于双目激光三维扫描成像和深度卷积神经网络融合的可靠缺陷定位和三维量化识别方法。.本项目通过理论和实验研究，开发了基于双目激光扫描三维成像与深度学习融合的连铸坯缺陷检测软硬件系统，并申报了相应专利3项，软件注册权1项，发表SCI检索论文3篇，EI检索论文2篇。.本项目按照原研究计划，先后开展了高温高噪声下，连铸坯表面缺陷检测三维量化检测方法研究。首次提出了基于深度学习技术的三维缺陷形态精确识别方法，以端到端的学习机制对系统架构进行了设计，特别针对工业产品表面缺陷检测，设计了深度卷积神经网络模型，同时，先后开发了两代实验室测试系统，通过课题研究，培养了5明硕士研究生。该项目目前已完成顶层设计到算法实现，为今后企业应用和推广奠定了理论基础和技术支撑。