【作者】 岳洪伟；

【导师】 汪仁煌；

【作者基本信息】 广东工业大学 ， 控制理论与控制工程， 2013， 博士

【摘要】 羽毛片作为羽毛球生产的主要原料,其质量直接影响到成品羽毛球的品质。羽毛杆有折痕会使得角质层受损从而降低耐打度。我国羽毛球产业仍然依赖密集人工完成折痕检测,检测速度和检测质量受个人主观因素影响很大。当然高强度的检测也对检测者视力造成伤害。因此利用机器视觉进行目标采集,通过特征提取完成折痕缺陷自动检测有着必要性,具有重要的工程实际意义和经济效益。毛杆的细长结构和拱度弯度加剧了光照不均,再加上折痕和背景界限模糊,这些因素都对折痕特征识别造成干扰。本文在国内外相关研究进展的基础上,对实验装置、图像分割提取、去噪增强、灰度校正以及折痕特征的度量、分类等多个方面进行了深入的研究,为羽毛球产业从人工密集型向高新科技型转变进行了有益的探索。本文完成的主要研究工作和成果如下：(1)毛杆边缘处理。由于羽毛杆具有较大的长宽比和不同的拱度弯度,这使得毛杆边缘噪声造成较大干扰,会增加对毛杆折痕的误判与错判的潜在可能。因此需要消除羽毛杆边缘残余绒毛以减少干扰。利用Ncut分割获得羽毛杆的初始轮廓,提出窄带单向膨胀方法完成初始轮廓的局部修正。为了改善过度分割,利用羽毛杆左右端顶点作为约束种子；然后将轮廓线两侧膨胀改为单向外侧膨胀,改善相邻轮廓线邻域重叠问题,完成初始轮廓的局部修正。实验结果表明该方法具有无需人工干预能够准确有效的实现羽毛杆分割。(2)经过对毛杆折痕统计分析表明大部分毛杆是无折痕的,有折毛杆往往也只有一到两个折痕。如果对二维图像遍历搜索缺陷会增加大量好的羽毛杆被判定为缺陷毛杆的潜在可能。这就要求检测方法在减少非折痕误测率基础上提高折痕检测。针对这一特点,本文首先将二维毛杆图像简化为一维灰度均值信号,利用小波变换的局部模极大值和信号奇异点之间的关系完成奇异点位置检测,确定缺陷存在位置,并完成缺陷子图像分割提取。(3)光线校正。侧光照明加剧了光照不均现象,再加上毛杆不同的弯度和拱度,这就难以使用校准模板进行光线校正。毛杆特有结构使得光线的衰减可以看作是个缓慢的过程。在此引入基线漂移概念,通过小波变换对灰度均值变化曲线进行多尺度分解,利用分解后所得的逼近信号充分逼近光线的变化,从而完成光线校正。(4)图像降噪。子图像分割后需要进行判断是否存在折痕缺陷。由于微小折痕容易受到噪声干扰,对折痕识别造成不利影响,为了改善图像质量,降低图像中存在的大量随机分布噪声的影响,以流形上的热导方程理论为工具对含噪数据流形上的概率测度进行了分析,通过引入热核并对流形上热核唯一性和正则性进行讨论,得到了羽毛杆含噪数据的密度函数与干净数据上的概率密度之间的对应关系,并设计了基于图的热导方程去噪算法。(5)通过对各种表面折痕缺陷的灰度特征、形状特征、分布特征的研究,提出一种改进的Radon变换用于毛杆折痕表面缺陷的识别算法。引入局部投影技术消除羽毛杆生理纹理干扰,通过改变尺度因子获得矩不变量矩阵,并采用奇异值分解获得特征不变量用于分类识别,从而在灰度域和梯度域得到两个单模态识别结果；最后采用决策级融合方法得到双模态融合识别结果,提高了非折痕识别率。(6)基于黎曼流形的毛杆折痕识别。利用协方差描述子对图像特征进行描述,重点研究了Log-Euclidean黎曼度量下协方差矩阵的李群结构,推导了黎曼流形中内积空间的度量形式,给出流形核函数表达；在仿射不变度量下讨论了流形上类间散度和类内散度计算,利用黎曼指数映射得到了样本的最佳映射空间,从而实现非线性空间的标签判别。在改进的局部保持投影算法中,运用流形核函数与局部保持投影相结合的方法进行毛杆特征提取：首先基于区域图像构造协方差矩阵作为图像特征,利用仿射不变度量作为样本点的距离测度,研究了流形上近邻点的选择以及包含数据类别信息的核矩阵构建；最后利用改进的局部保持投影算法对毛杆图像进行降维。(7)提出黎曼流形上的核稀疏表示方法进行毛杆折痕识别。首先利用Bregman散度作为黎曼流形上的距离度量,并以此构造了流形上的核函数,把协方差矩阵映射到再生核希尔伯特空间。通过稀疏表示理论建立了字典学习的数学模型,结合凸理论给出了黎曼流形上的核稀疏最大化的字典学习算法。最后利用SRC分类器完成测试样本鉴别。实验结果证实了所提方法的有效性。论文最后阐述了研究的创新点及主要研究成果、展望。更多还原

【Abstract】 Feather as the main raw material of badminton production, its quality directly affects the quality of the finished product badminton.Creases of feather quill occur when the cuticle is damaged and the fibres of the cortex unravel.However, the quality testing of badminton industry relies a lot on manual work in our country, which results in that the detecting speed and quality are influenced by personal. Actually high strength testing is harmful to eyesight of the workers. Consequently, it’s of great engineering significances and potential economic effect to study crease defect automatic detection by feature extraction using machine vision for testing image acquisition.Due to slender structure of feather quill with variable width, camber and curvature, in addition, boundary between crease of feather quill and background is fuzzy, all of these factors cause disturbance to the crease feature recognition.Based on detailed research development from home and abroad, this thesis explores experimental device,image segmentation,denoising,graylevel correction and crease features classification, etc.,deeply.There are some helpful researches about badminton industry from labour-intensive industry to technology-intensive industry.The related work and harvest in this paper as follows:(1) Image edge processing. Due to slender structure of feather quill with variable width, camber and curvature, this makes marginal noise caused disturbance to the crease feature recognition to increase the potential for misunderstanding and miscalculation. So need to eliminate the residual edge villi of feather quill, edge to reduce noise.The method partitions a feather image into clusters with normalized cut method to obtain an initial contour of the feather quill. The proposed narrow unidirectional expansion method with a region-scalable fitting term is used to adjust the initial boundary for the final result. This paper presents a method of image segmentation by combining certain hard constraints for segment by indicating certain seeds with unidirectional expansion. The method utilizes left and right vertex of feather quill as seeds which are passed by active contour to better over segmentation; changes bidirection dilation to inside direction dilation to improve the overlap of adjacent contour neighborhoods and reduce the computation scale. Experiment results show that the proposed method without human intervention can effectively realize the image segmentation effectively.(2) Statistical analysis of feather quill crease shows that most of feather quill is no crease or some feather quill only have one or two crease. Exhaustive searches of two-dimensional image can increase the potential for misunderstanding and miscalculation. As most of feather quill without creases in actual production,detection method requires not only effective feature extraction, but also a high degree of non-crease detection accuracy. Aiming at this characteristic, a detection method of feather quill crease is proposed. The feather quill image is transformed into one dimensional signal, using the relationship between wavelet transform modulus maxima and singular point,the crease coordinate can be prejudged. Subimage can be extracted through the coordinate to reduce misjudgement caused by image traversal.(3)Light correction.Side lighting exacerbated the uneven illumination phenomenon, combined with different camber and camber feather quill, so it is difficult to use correction calibration template to correction light. Due to the unique structure of feather quill, the attenuation of light can be seen as a slow process. The concept of "baseline drift" is introduced,and multi-scale decompositions are carried out for mean gray level curve. Using wavelet approximation coefficients to close characteristic of baseline drift completes light correction.(4) Image denoising.Tiny creases of feather quill is easily disturbed by noise that causing adverse effects on recognition performance. For the shortcomings of the traditional noise reduction algorithm in the process of image, i.e., it is sensitive to the noise resulting in the problem of weakened creases characteristics, this paper discusses probability measure of noisy image on manifolds based on heat equation theory, analyzes regularity and uniqueness of heat kernels on manifolds,and gets the corresponding relationship between density function of noisy data and probability density of clean data. A denoising algorithm of feather quill based on the heat equation of graph is proposed. In the method, de-noised image can be obtained through iterative solution of equation based on feather quill image representing as an undirected weighted graph. The Experimental results show that this suggested method can get better effect comparing with other transform domain algorithms. (5) Appropriate shape description method. According to the surface defect characteristics such as shape, size and gray distribution,a novel method based on improved Radon transform is proposed. In order to solve the scaling and translation sensitivity of Radon transformation, an improved Radon transformation is used to extract moment invariants of target region and introduces local projection technology to eliminate interference physiological texture of feather quill. Obtaining invariants matrix by changing the scale factor, singular value decomposition is provided here to obtain feature invariant for classification and recognition. Finally,the final recognition result of the system is achieved by the fusion of identification results of the gray domain and gradient domain at the decision level to overcome the limitations of single-modal and reduce the misjudgment of non-crease effectively.(6) Feather quill crease detection method based on Riemannian manifold.Research is mainly focused on the covariance matrix lie group structure with Log-Euclidean metric with double invariance properties and manifold kernel function expression with covariance matrices as the crease descriptors of feather quill;then this article designs algorithm with metrics of inner product space and manifold kernel function expression which are deduced. An affine invariance metric which is adopted to make the space meet the requirement of Riemannian manifold is used to adjust class variance and within class variance. Finally, in order to implement the discrimination in nonlinear space, the best projection space of samples is gotten using the Riemannian index mapping. A feather quill crease recognition method based on locality preserving projection and manifold kernel function is proposed for feature extraction. Firstly, covariance matrices are computed as the crease descriptors of feather quill, and an affine invariance metric which is adopted to make the space meet the requirement of Riemannian manifold is used to measure the distance between the two samples. Secondly, the neighbors of a selected point can be determined by the proposed manifold kernel function to make choice of the nearest neighboring points in line with the hypothesis of data distribution with non-linear manifold. The kernel matrix is defined based on the manifold distance and category labels. Finally, the locality preserving projections algorithm is used to reduce the dimensionality of the feather quill images.(7) A feather quill crease recognition method based on sparse representation on Riemannian manifold is proposed for feature recognition. Firstly, Bregman divergence which is adopted to make the space meet the requirement of Riemannian manifold is used to measure the distance between the two samples, and the kernel function on Riemannian manifold is constructed. Secondly, the sample datasets are mapped into the reproducing kernel Hilbert space by kernel method, and kernel sparse representation coefficient is obtained. Finally, a mathematical model of dictionary learning is constructed and an efficient algorithm is proposed for dictionary learning according to the convex theory. The simulated experimental results verify the effectiveness of the proposed method which achieves better performance compared with many popular recognition algorithms.At the end of this dissertation, the main research is summarized. It makes out the main innovations and research achievements, and also points out the problems and issues which need to further research.更多还原