【作者】 崔扬；

【导师】 周泽魁；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2004， 博士

【摘要】 在生产过程中，为了有效利用原材料和避免次品混入成品中，常常需要对原材料和产品进行检测。但长期来，这些检测过程大多靠手工实现，大大降低了生产效率，增加了生产成本，随着计算机视觉技术的发展，原来靠手工实现的检测过程现在都可以由计算机来实现，从而将工人从繁重的工作中解放出来，降低了生产成本。皮革是制鞋的主要原材料，制鞋的主要工序是在优质皮革上排放和切割各种鞋样部件，但是皮革表面不可避免的存在各种缺陷，如虫咬、疤痕和划伤等，为了有效利用原料，在排样和切割之前要首先定位皮革表面的缺陷区域。同时由于所要生产的皮鞋款式不尽相同，需要在皮革正面轧制各种花纹，因此皮革具有典型的纹理表面。本文基于图像处理技术，结合皮革的特点，提出了一系列具有针对性的算法，实现了对皮革表面缺陷的自动化检测。 本文基于Fisher准则和神经网络设计了两种分类器实现了对皮革正反面的自动分类。由于排样和切割在皮革正面进行，从而缺陷检测也主要在皮革正面进行，因此自动化缺陷检测系统首先要自动识别皮革正反面。由于皮革正面和反面具有明显不同的纹理特征，本文用共生矩阵提取皮革正反面的纹理特征，并根据提取的特征向量采用分类器对皮革正反面进行自动识别和分类。 为了有效去除与纹理信息同处于高频的噪声，基于小波包分解提出了一种新颖的去噪方法。图像的质量对图像检测能否顺利进行影响较大，在皮革图像的获取过程中，由于自然环境和系统自身的影响，采集到的皮革图像中掺杂有大量的噪声，因此在缺陷检测前，需要采用适当的去噪方法提高图像的信噪比。由于皮革图像中的纹理和噪声都属于高频信息，因此常用的去噪方法不能有效去除图像中的噪声，鉴于信号和噪声在小波分解的过程中呈现出不同特性，提出了一种新颖的去噪算法，首先对皮革图像进行小波包分解，并在各个子带图像中对信号和噪声采用不同的处理方法，该算法既能有效去除图像中的大量噪声，又保留了图像中的纹理和缺陷信息。 本文基于小波分解提出了一种有效的皮革图像融合算法。由于光线和图像采集设备视角的限制，对皮革采集的一幅图像可能无法体现出皮革的整体信息，必须对采集到的多幅互补和冗余图像进行融合来获得皮革的整体特征。基于小波分解，提出了一种融合算法，该算法在小波分解的子带图像中提取局部边缘信息作为融合规则，较好的实现了两幅互补皮革图像的融合。 由于皮革表面的缺陷信息无法预先得知，本文提出了一种基于邻域信息的改进模糊C-均值聚类方法。该方法首先采用共生矩阵提取皮革的纹理特征向量，然后用改进的模糊C-均值算法对特征向量进行聚类分析，该算法可以准确的检测出皮革表面的缺陷信息。同时为了自适应确定纹理分割区域数，引入了一种基于类内方差和类间方差的有效值函数。更多还原

【Abstract】 In order to utilize materials effectively and avoid mixing inferior products with good products in production, it is necessary to detect materials and products. For a long time the detection of materials and products is done manually, this reduces production efficiency and adds costs. As the result of progress of computer vision, computer can make the detection now, this liberates workers from heavy labor and reduces costs. Leather is main material for making shoes, in shoemaking the main working procedures are nesting and cutting all kinds of shoe parts on leather with high quality, but it is inevitable that there are all kinds of defects on leather, such as scratch, mites (insect bites) and scars, so it is necessary to detect the positions and size of defects before nesting and cutting shoe parts in order to utilize materials effectively. Because of printing all kinds of pattern on the face of leather depending on style of shoes to be made, leather has typical texture on the face. Based on image processing this paper presented a series of algorithms aimed at leather which detected defects on the face of leather automatically.This paper presented two classifiers based on Fisher criterion and Nerve Net (NN) to classify the face and the inverse of leather automatically. Because of nesting and cutting being done on the face of leather, so detecting defects is done on the face of leather and it is important for automatic system of detecting defects to recognize the face and the inverse of leather before nesting and cutting shoe parts. Based on different texture of the face and the inverse of leather, this paper extracted characteristic vector of the face and the inverse of leather using co-occurrence matrix and classified the face and the inverse of leather using Fisher classifier and NN base on characteristic vector.This paper presented a novel denoising algorithm based on wavelet packet decomposition in order to wipe off noise mixed with texture lying in high frequency. The quality of image has a great effect on image detection, while acquiring leather image, because of the effect of environment and system, a lot of noise is mixed with leather image, and so it is necessary to improve SNR by denoising image using adaptive methods before detecting defects. Because both texture and noise lie in high frequency, so usual methods for denoising cannot denoise leather image effectively, based on the different character signal and noise show as wavelet decomposition goes on, this paper presented a novel denoising algorithm which decomposed the leather image using wavelet packet and treated signal and noise with different methods in subband images, this algorithm not only can wipe off a lot of noise in leather image, but also can keep the texture and defects in leather image.This paper presented an effective fusion algorithm for leather images based on wavelet decomposition. Due to the limitation of light source and viewing angle of device for acquiring image, only one image of leather cannot exhibit the whole character of leather, it is necessary to fuse many images with redundancy and complementarity to get the whole character of leather. Based on wavelet decomposition this paper presented a fusion algorithm, which extracted the local edge info in subband images as fusion rule and fused two leather images with complementarity.For not getting the prior info of defects on the face of leather, this paper presented a modified FCM based on neighbor info. This algorithm first extracted the texture characteristic vectors ofleather using co-occurrence matrix, then clustered the characteristic vectors using modified FCM, this algorithm can detect the defects on leather exactly. In order to decide the number of regions adaptively, this paper introduced a valid function based on intra-class and inter-class variance.更多还原