车辆皮革瑕疵智能检测方法研究

发布时间：2018-03-06 20:23

本文选题：皮革　切入点：瑕疵检测　出处：《重庆理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着生活水平的提高,私家车保有量激增,消费者在关注性能同时也开始注重内部饰品质量。皮革作为其内部座椅等主要器件的重要材料,其品质被严格要求,但由于牛皮等皮革原材料在生长生产过程中的蚊虫叮咬及人为误伤,使其表面不可避免的存在各种瑕疵,因此需要定位其表面瑕疵部分,以便控制产品品质和指导后续生产加工。目前汽车座椅生产厂家采用的人工皮革瑕疵查找方式存在误检率高,效率低等缺点,而基于计算机视觉方式检测的可行性及优势使得其希望能引入计算机视觉方式,替代人工,使检测更安全,更效率,更稳定,更客观同时也更节约成本。针对目前皮革瑕疵检测中瑕疵与非瑕疵区域之间的低对比度和复杂随机纹理的干扰等原因造成其检测难度大、速度慢;且没有瑕疵检测效果的客观量化评判方法等问题,通过初步分析皮革瑕疵样本,最终对人工视觉查找有难度的皮革微小瑕疵的自动检测作如下几方面工作:首先,建立皮革瑕疵检测算法的评价方法。将瑕疵检测看作特殊的分类工作,参考文本分类的评价指标,提出一种基于召回率和准确率的评价体系。通过一种人工画笔标记样本图像中瑕疵区域并经过识别、分割处理的方式获取黄金标准,并将其作为算法检测结果的参考。然后通过计算定义的P、R、f及综合评价参数F等评价指标,实现对瑕疵检测结果像素级的数字化评价,为算法研究提供客观的指导。其次,基于皮革瑕疵查找是人眼注意选择机制的一种表现,提出基于视觉显著度的瑕疵检测模型。基于该模型的瑕疵检测算法,首先提取颜色、亮度特征,利用图像本身作为模板进行对比计算显著度图;然后根据随机均匀分布纹理图像中“突出”部分显著度高的特征,通过最显著像素点利用区域增长法分割定位瑕疵区域。分别利用该算法与阈值法、基于模糊聚类及支持向量机等现有瑕疵检测算法对皮革瑕疵样本对比实验,结果表明该算法解决了模板和有效特征提取困难的问题,能有效检测微小皮革瑕疵。最后,对纹理分析技术进行研究,设计了一种基于灰度共生矩阵的纹理表述,结合基于视觉显著度的皮革瑕疵检测算法以进一步提高检测效果。首先对图像进行灰度共生矩阵的统计,然后计算每个像素的灰度分布频率作为其纹理表示。该纹理特征较传统基于共生矩阵的能量等纹理特征计算量大、速度慢等劣势更适用于应用在有一定实时性要求的皮革瑕疵检测中,实验结果也表明结合纹理分析后在检测效果也有一定提高,具有一定意义。
[Abstract]:With the improvement of living standards and the increase of private car ownership, consumers are paying more attention to the quality of interior ornaments. Leather, as an important material for its internal seats and other main devices, has been strictly required for its quality. However, due to mosquito bites and manmade injuries in the course of growth and production of leather raw materials such as cowhide, there are inevitably various defects on its surface, so it is necessary to locate the defective parts of the surface. In order to control the product quality and guide the subsequent production and processing. At present, the artificial leather flaw detection method adopted by the automobile seat manufacturers has the disadvantages of high false detection rate and low efficiency. Based on the feasibility and advantages of computer vision detection, it hopes to introduce computer vision instead of manual, and make the detection safer, more efficient and more stable. It is more objective and cost saving. It is difficult and slow to detect because of the low contrast and the interference of complex random texture between the defect and the non-defect area in the current leather flaw detection. And there is no objective quantitative evaluation method of defect detection effect. Through the preliminary analysis of leather defect samples, the automatic detection of the difficult small leather defects in artificial vision is done as follows: first of all, The evaluation method of leather defect detection algorithm is established. The defect detection is regarded as a special classification work, and the evaluation index of text classification is referred to. This paper presents an evaluation system based on recall and accuracy. It is used as the reference of the algorithm detection result. Then, the digital evaluation of the pixel level of the defect detection result is realized by calculating the defined PGR f and the comprehensive evaluation parameter F, which provides objective guidance for the research of the algorithm. This paper presents a defect detection model based on visual saliency, based on which the color and luminance features are first extracted. Using the image itself as a template to compare and calculate the saliency map, and then according to the feature of high salience in the "prominent" part of the randomly distributed texture image, The most significant pixel points are segmented by using region growth method to locate the defect region. Using this algorithm and the threshold method respectively based on the existing defect detection algorithms such as fuzzy clustering and support vector machine the contrast experiment of leather defect samples is carried out. The results show that the algorithm solves the difficult problem of template and effective feature extraction, and can effectively detect small leather defects. Finally, a texture representation based on gray level co-occurrence matrix is designed, which is based on the research of texture analysis technology. Combined with the leather defect detection algorithm based on visual saliency to further improve the detection effect. Firstly, the gray level co-occurrence matrix of the image is counted. Then the gray distribution frequency of each pixel is calculated as its texture representation, which is more expensive than the traditional energy and other texture features based on co-occurrence matrix. The disadvantages such as slow speed and so on are more suitable for the leather defect detection with certain real-time requirements. The experimental results also show that the detection effect is also improved with texture analysis, which has a certain significance.
【学位授予单位】：重庆理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U466;TP391.41

【相似文献】