CN103926255A

Movatterモバイル変換

Info

Publication number: CN103926255A
Application number: CN201410173860.5A
Authority: CN
Inventors: 白瑞林; 何薇; 吉峰; 李新
Original assignee: XINJE ELECTRONIC CO Ltd; Jiangnan University
Current assignee: XINJE ELECTRONIC CO Ltd; Jiangnan University
Priority date: 2014-04-26
Filing date: 2014-04-26
Publication date: 2014-07-16

Abstract

本发明提供了一种布匹表面瑕疵的在线视觉检测方法，其特征是通过Gabor滤波器和小波神经网络的结合，有效提取出布匹表面纹理的宽度方向等信息，能够对同种类布匹训练求取最优解后构建Gabor滤波器进行在线实时检测，保证了在线检测的速度与精度；针对不同种类的瑕疵，分别利用奇对称、偶对称Gabor滤波器保证块状瑕疵与线状瑕疵均能够准确、高效地检测出来。在用线阵相机高速实时采集图像的条件下，能够有效的提高检测速度，降低漏检和误检率。

The invention provides an online visual detection method for cloth surface defects, which is characterized in that information such as the width direction of the cloth surface texture can be effectively extracted through the combination of the Gabor filter and the wavelet neural network, and can obtain the best results for the same type of cloth training. After optimization, build a Gabor filter for online real-time detection, which ensures the speed and accuracy of online detection; for different types of defects, use odd-symmetric and even-symmetric Gabor filters to ensure that both blocky and linear defects are accurate and efficient detected out. Under the condition of high-speed real-time image acquisition with line array cameras, the detection speed can be effectively improved, and the missed detection and false detection rates can be reduced.

Description

Translated fromChinese

一种基于小波神经网络的布匹表面瑕疵检测方法A Method for Detection of Cloth Surface Flaws Based on Wavelet Neural Network

技术领域technical field

本发明涉及一种基于机器视觉的布匹瑕疵实时视觉检测方法，具体是指一种在线阵光源下，通过线阵相机对工业现场中高速传送的布匹表面瑕疵进行检测并即时记录的图像检测方法。The invention relates to a real-time visual detection method for cloth defects based on machine vision, specifically an image detection method for detecting and recording immediately the surface defects of cloth conveyed at high speed in an industrial site through a line array camera under a line array light source.

背景技术Background technique

工业生产过程里，随着技术水平的不断提高，市场对产品质量的要求也一再提升。在纺织行业中，布匹的质量检测要求随着这种发展趋势愈加严格。但因为纺织品产量持续增大，生产线工业化水准提升，传统的人工检测法已经跟不上自动化发展的速度，受制于检查人员的主观因素以及精神状态，并且存在着人工检验速度慢、成本高、标准化程度低、误检率大等劣势，快速精确地检测出纺织品瑕疵成为生产过程中亟待解决的问题。In the industrial production process, with the continuous improvement of the technical level, the market's requirements for product quality have also been continuously improved. In the textile industry, the quality inspection requirements of cloth are becoming more and more strict with this development trend. However, due to the continuous increase in the output of textiles and the improvement of the industrialization level of the production line, the traditional manual inspection method has been unable to keep up with the speed of automation development. Due to the disadvantages of low degree and high false detection rate, it is an urgent problem to be solved in the production process to quickly and accurately detect textile defects.

面对这样的的需求，国外的一些大型企业在工业上已经有了一定规模的应用，主要代表产品有以色列EVS公司的IQ-TEX4自动在线检测系统，美国BMS公司的Cyclops自动在线织物检测系统等，但成本高昂、维护不易，在国内并不普遍推广适用。目前，研究者主要采用基于统计学方法、频域变换法、模型法等方法对布匹图像进行处理，以求准确检测到瑕疵，由于布匹表面带有纹理干扰，瑕疵种类繁复，正确地提取出瑕疵区域成为布匹表面检测中的重点和难点。Faced with such a demand, some large-scale foreign enterprises have already had a certain scale of application in the industry. The main representative products include the IQ-TEX4 automatic online inspection system of the Israeli EVS company, the Cyclops automatic online fabric inspection system of the American BMS company, etc. , but the cost is high and maintenance is not easy, so it is not widely applied in China. At present, researchers mainly use methods based on statistical methods, frequency domain transformation methods, and model methods to process cloth images in order to accurately detect defects. Due to the texture interference on the cloth surface and the variety of defects, it is necessary to correctly extract the defects. The area becomes the focus and difficulty in the detection of the cloth surface.

由于在检测过程中，出布速度快，布匹幅面较大，检测精度要求高，选用高分辨率并适用于高速采集过程的的线阵相机作为图像采集传感器已经越来越成为主流的检测方式。Due to the fast cloth output speed, large cloth size, and high detection accuracy requirements during the detection process, it has become more and more mainstream to use a high-resolution linear array camera suitable for high-speed collection as an image acquisition sensor.

发明内容Contents of the invention

本发明目的在于提出一种通用性强的基于机器视觉的布匹瑕疵检测方法，取代传统的效率低下的人工检测法。The purpose of the present invention is to propose a machine vision-based cloth defect detection method with strong versatility to replace the traditional low-efficiency manual detection method.

针对这个目的，本发明通过如下技术方案实现：For this purpose, the present invention realizes through following technical scheme:

离线状态：Offline status:

(1)利用线阵相机实时获取无瑕疵的布匹图像，调节布匹的传送速度、相机采集频率以及相机光圈焦距等参数，实时获得无瑕疵的布匹图像序列作为样本。(1) Use the line array camera to acquire flawless cloth images in real time, adjust the cloth transmission speed, camera acquisition frequency, camera aperture focal length and other parameters, and obtain the flawless cloth image sequences as samples in real time.

(2)对获取的样本图像进行中值滤波处理以抑制噪声去除干扰点，利用直方图均衡化增强图像的对比度以凸显纹理。(2) Perform median filtering on the acquired sample image to suppress noise and remove interference points, and use histogram equalization to enhance the contrast of the image to highlight the texture.

(3)构建一个三层前馈神经网络结构，采用一个虚部Gabor小波作为隐层的激励函数，构建参数向量组。(3) Construct a three-layer feed-forward neural network structure, use an imaginary part Gabor wavelet as the activation function of the hidden layer, and construct a parameter vector group.

(4)利用Levenberg-Marquardt(LM算法)针对每个Gabor小波求解出最优参数，最终得到相互对应的奇对称Gabor滤波器组和偶对称Gabor滤波器组。(4) Use the Levenberg-Marquardt (LM algorithm) to solve the optimal parameters for each Gabor wavelet, and finally get the corresponding odd symmetric Gabor filter bank and even symmetric Gabor filter bank.

在线状态：online status:

(1)保持离线状态的相机参数，实时获取待测布匹图像。通过中值滤波处理去除噪声干扰点，利用直方图均衡化凸显布匹图像的纹理，每个采样周期内采集布匹图像形成图像序列进行检测。(1) Keep the camera parameters in the offline state, and obtain the image of the cloth to be tested in real time. The noise interference points are removed by median filtering, the texture of the cloth image is highlighted by histogram equalization, and the cloth image is collected in each sampling period to form an image sequence for detection.

(2)分别用离线训练时得到的奇对称Gabor滤波器组和偶对称Gabor滤波器组对待检测图片滤波处理，检测块状与线状瑕疵。(2) Use the odd symmetric Gabor filter bank and the even symmetric Gabor filter bank obtained during offline training to filter the image to be detected to detect blocky and linear defects.

(3)对得到的滤波结果进行融合处理，并对融合图像平滑滤波以及二值化，最终得到瑕疵区域。(3) Fusion processing is performed on the obtained filtering results, and smooth filtering and binarization are performed on the fusion image to finally obtain the defect area.

本发明的有益效果是：本发明提供了一种基于Gabor滤波器和小波神经网络的图像瑕疵检测方法，对于布匹图像的纹理信息能够有效的提取，并利用离线训练缩短了算法的执行时间，构建的奇对称Gabor滤波器可以很好的检测块状瑕疵，而偶对称Gabor滤波器则在检测线性边缘时有着很大的优势，面对不同类别的瑕疵能够保证检测成功率。在用线阵相机高速实时采集图像的条件下，能够有效的提高检测速度，降低漏检和误检率。The beneficial effect of the present invention is: the present invention provides a kind of image blemish detection method based on Gabor filter and wavelet neural network, can effectively extract the texture information of cloth image, and utilize off-line training to shorten the execution time of the algorithm, construct The odd symmetric Gabor filter can detect blocky defects very well, while the even symmetric Gabor filter has a great advantage in detecting linear edges, and can guarantee the detection success rate in the face of different types of defects. Under the condition of high-speed real-time image acquisition with line array cameras, the detection speed can be effectively improved, and the missed detection and false detection rates can be reduced.

附图说明Description of drawings

图1本发明的整体系统构建图Overall system construction diagram of the present invention of Fig. 1

图2本发明算法整体流程图Fig. 2 overall flowchart of the algorithm of the present invention

具体实施方式Detailed ways

为使本发明的目的、技术方案和优点等更加清楚明白，以下结合具体实施例，并参照附图，对本发明进一步详细说明。In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

本发明的基本目的是检验布匹的表面瑕疵，分为离线训练过程及在线检测过程，装置的硬件构建如图1所示，算法整体流程如图2所示。离线训练过程中，通过对无瑕疵样本图像进行小波网络算法处理，并利用LM算法迭代寻优得到最优参数组，构建出对应奇对称Gabor滤波器组和偶对称Gabor滤波器组。在线检测过程中利用得到的滤波器组对待检测图像进行滤波处理，再图像融合，平滑滤波处理后最终得到瑕疵区域。The basic purpose of the present invention is to test the surface flaws of the cloth, which is divided into an offline training process and an online detection process. The hardware construction of the device is shown in Figure 1, and the overall algorithm flow is shown in Figure 2. In the offline training process, the wavelet network algorithm is used to process the flawless sample images, and the optimal parameter set is obtained by using the LM algorithm to iteratively optimize, and the corresponding odd symmetric Gabor filter bank and even symmetric Gabor filter bank are constructed. In the online detection process, the obtained filter bank is used to filter the image to be detected, and then the image is fused, and the defect area is finally obtained after smoothing and filtering.

进一步的，离线训练过程具体实现步骤为：Further, the specific implementation steps of the offline training process are:

(1)根据所要求的验布精度和速度，设定布匹传送速度，相机采样频率，调整相机位置等，采集一系列图像序列作为样本图像。(1) According to the required cloth inspection accuracy and speed, set the cloth transmission speed, camera sampling frequency, adjust the camera position, etc., and collect a series of image sequences as sample images.

(2.1)对得到的样本图像利用中值滤波做去噪处理。(2.1) Use median filter to denoise the obtained sample image.

(2.2)对滤波后图像做直方图均衡化处理。(2.2) Perform histogram equalization processing on the filtered image.

在一幅图像中灰度级r_k出现的概率近似为：The probability of gray level r_k appearing in an image is approximately:

$P_{r} (r_{k}) = \frac{n_{k}}{n}$ k＝0，1，2，...，L-1 $P_{r} (r_{k}) = \frac{{no}_{k}}{no}$ k=0, 1, 2, ..., L-1

其中n是图像的像素和，n_k是图像中灰度级为r_k的像素个数，L是图像中可能的灰度级总数。则有输出灰度级sk的变换函数为：where n is the sum of pixels in the image,_nk is the number of pixels in the image whose gray level is r_k , and L is the total number of possible gray levels in the image. Then the transformation function of the output gray level sk is:

$s_{k} = T (r_{k}) = Σ_{j = 0}^{k} P_{r} (r_{j}) = Σ_{j = 0}^{k} \frac{n_{j}}{n}$ k＝0，1，2，...，L-1 ${the s}_{k} = T (r_{k}) = Σ_{j = 0}^{k} P_{r} (r_{j}) = Σ_{j = 0}^{k} \frac{{no}_{j}}{no}$ k=0, 1, 2, ..., L-1

通过该变换函数能将输入图像中灰度级为r_k的各个像素映射到输出图像中灰度级为s_k的对应像素。Through this transformation function, each pixel whose gray level is r_k in the input image can be mapped to the corresponding pixel whose gray level is s_k in the output image.

(3)构建小波神经网络。(3) Construct wavelet neural network.

构建一个三层前馈神经网络结构。令一副灰度图像为f(x，y)，其中(x，y)为像素的位置标号，f为对应的像素值，使用Gabor小波网络逼近该图像则有如下表示形式： $\hat{f} (x, y) = Σ_{i = 1}^{N} w_{i} g_{0}^{i} (x, y) + \overset{&OverBar;}{f} .$ Build a three-layer feed-forward neural network structure. Let a grayscale image be f(x, y), where (x, y) is the position label of the pixel, f is the corresponding pixel value, and the Gabor wavelet network is used to approximate the image as follows: $\hat{f} (x, the y) = Σ_{i = 1}^{N} w_{i} g_{0}^{i} (x, the y) + \overset{&OverBar;}{f} .$

其中，wi为第i个隐层节点到输出节点之间的权重Among them, wi is the weight between the i-th hidden layer node and the output node

$\begin{matrix} {g g}_{00}^{i i} ((x x,, y the y)) = = exp exp ((- - \frac{11}{22} {{{[[\frac{((x x - - {t t}_{x x}^{i i})) cos cos {θ θ}^{i i} - - ((y the y - - {t t}_{y the y}^{i i} sin sin {θ θ}^{i i}))}{{σ σ}_{x x}^{i i}}]]}^{22} + + {[[\frac{((x x - - {t t}_{x x}^{i i})) sin sin {θ θ}^{i i} - - ((y the y - - {t t}_{y the y}^{i i})) cos cos {θ θ}^{i i}}{{σ σ}_{y the y}^{i i}}]]}^{22}}})) \\ \times \times sin sin ((22 π π {ω ω}_{x x}^{i i} [[((x x - - {t t}_{x x}^{i i})) cos cos {θ θ}^{i i} - - ((y the y - - {t t}_{y the y}^{i i})) sin sin {θ θ}^{i i}]])) \end{matrix}$

和为坐标轴上的平移参数，和为第i个隐层节点Gabor小波的径向频率带宽，θⁱ为第i个Gabor小波的旋转角度，为中心频率。 and is the translation parameter on the coordinate axis, and is the radial frequency bandwidth of the i-th hidden layer node Gabor wavelet, θⁱ is the rotation angle of the i-th Gabor wavelet, is the center frequency.

这些参数构成参数向量 $η = (t_{x}^{i}, t_{y}^{i}, θ^{i}, σ_{x}^{i}, σ_{y}^{i}, ω_{x}^{i}, w_{i}) .$ These parameters form the parameter vector $η = (t_{x}^{i}, t_{the y}^{i}, θ^{i}, σ_{x}^{i}, σ_{the y}^{i}, ω_{x}^{i}, w_{i}) .$

(4)利用LM算法得到最优参数组，得到最优Gabor滤波器组。(4) Using LM algorithm to obtain the optimal parameter group, and obtain the optimal Gabor filter group.

(4.1)将得到的参数向量组进行寻优，整个网络的训练过程可以表达为寻优就是使能量函数最小化的过程。(4.1) Optimizing the obtained parameter vector group, the training process of the whole network can be expressed as Optimizing is the process of minimizing the energy function.

由于Gabor基函数非正交，需要采用LM算法进行迭代训练得到最优参数。Since the Gabor basis functions are non-orthogonal, it is necessary to use the LM algorithm for iterative training to obtain the optimal parameters.

步骤如下：Proceed as follows:

N＝1，使用LM算法找到最优值 $η_{1} = (t_{x}^{1}, t_{y}^{1}, θ^{1}, σ_{x}^{1}, σ_{y}^{1}, ω_{x}^{1}, w_{1})$ N=1, use LM algorithm to find the optimal value $η_{1} = (t_{x}^{1}, t_{the y}^{1}, θ^{1}, σ_{x}^{1}, σ_{the y}^{1}, ω_{x}^{1}, w_{1})$

N＝2，在不变的前提下，用LM算法找到最优值 $η_{2} = (t_{x}^{2}, t_{y}^{2}, θ^{2}, σ_{x}^{2}, σ_{y}^{2}, ω_{x}^{2}, w_{2})$ N=2, in Under the premise of being unchanged, use the LM algorithm to find the optimal value $η_{2} = (t_{x}^{2}, t_{the y}^{2}, θ^{2}, σ_{x}^{2}, σ_{the y}^{2}, ω_{x}^{2}, w_{2})$

i、N＝i，在保证η₁，η₂，...，η_i-1不变的前提下，用LM算法找到最优值i, N=i, under the premise of ensuring that η₁ , η₂ ,..., η_i-1 remain unchanged, use LM algorithm to find the optimal value

${η η}_{i i} = = (({t t}_{x x}^{i i},, {t t}_{y the y}^{i i},, {θ θ}^{i i},, {σ σ}_{x x}^{i i},, {σ σ}_{y the y}^{i i},, {ω ω}_{x x}^{i i},, {w w}_{i i}))$

(4.2)这样就得到一个偶对称的Gabor滤波器形式：(4.2) In this way, an even symmetric Gabor filter form is obtained:

${g g}_{even even} ((x x,, y the y)) = = \frac{11}{22 π π {σ σ}_{x x} {σ σ}_{y the y}} exp exp {{- - \frac{11}{22} [[{((\frac{x x}{{σ σ}_{x x}}))}^{22} + + {((\frac{x x}{{σ σ}_{y the y}}))}^{22}]]}} \times \times cos cos ((22 π π {ω ω}_{x x} x x))$

奇对称的Gabor滤波器形式：Odd symmetric Gabor filter form:

${g g}_{odd odd} ((x x,, y the y)) = = \frac{11}{22 π π {σ σ}_{x x} {σ σ}_{y the y}} exp exp {{- - \frac{11}{22} [[{((\frac{x x}{{σ σ}_{x x}}))}^{22} + + {((\frac{x x}{{σ σ}_{y the y}}))}^{22}]]}} \times \times sin sin ((22 π π {ω ω}_{x x} x x))$

在线检测过程具体实现步骤为：The specific implementation steps of the online detection process are as follows:

(1)保持离线状态的相机参数，实时获取待测布匹图像，每个采样周期内采集布匹图像形成图像序列进行检测(1) Keep the camera parameters in the offline state, obtain the image of the cloth to be tested in real time, and collect the image of the cloth in each sampling period to form an image sequence for detection

(2)通过中值滤波处理去除噪声干扰点，利用直方图均衡化凸显布匹图像的纹理。(2) The noise interference points are removed by median filtering, and the texture of the cloth image is highlighted by histogram equalization.

(3)对待测图像经过Gabor滤波器组滤波处理，将得到的各图像融合，得到瑕疵区域。(3) The image to be tested is filtered by the Gabor filter bank, and the obtained images are fused to obtain the defect area.

(3.1)将待测图像与离线步骤得到的Gabor滤波器组进行滤波处理：(3.1) Filter the image to be tested and the Gabor filter bank obtained in the offline step:

偶对称的Gabor滤波器形式：Even symmetric Gabor filter form:

奇对称的Gabor滤波器形式：Odd symmetric Gabor filter form:

(3.2)为了提高检测成功率，减少误判，需要对两个Gabor滤波器的滤波结果进行融合，就瑕疵检测而言，需要削弱背景纹理并加强瑕疵区域的响应，先将两个滤波器的输出结果(O_even(x，y)，O_odd(x，y))分别归一化处理：(3.2) In order to improve the detection success rate and reduce misjudgment, it is necessary to fuse the filtering results of two Gabor filters. In terms of defect detection, it is necessary to weaken the background texture and strengthen the response of the defect area. The output results (O_even (x, y), O_odd (x, y)) are normalized respectively:

${O o}^{' '} ((x x,, y the y)) = = \frac{O o ((x x,, y the y)) - - min min ((O o ((x x,, y the y))))}{max max ((O o ((x x,, y the y)))) - - min min ((O o ((x x,, y the y))))}$

(3.3)对图像进行融合：(3.3) Fusion of images:

F(x，y)＝O′_even(x，y)+O′_odd(x，y)-O′_even(x，y)×O′_odd(x，y)F(x, y)=O'_even (x, y)+O'_odd (x, y)-O'_even (x, y)×O'_odd (x, y)

对得到的图像去噪处理，最后分离得到凸显瑕疵的二值图像。The obtained image is denoised, and finally a binary image with highlighted defects is obtained by separation.

(4)如果此帧图像没有一处为瑕疵区域，则不保存此图像序列。继续检测下一帧图像，若是此帧图像出现瑕疵，保存此瑕疵图像与位置信息至结构体并继续检测下一幅图像。(4) If there is no defect area in the frame image, the image sequence will not be saved. Continue to detect the next frame of image. If there is a defect in this frame of image, save the defect image and location information to the structure and continue to detect the next image.