基于卷积神经网络的印刷电路板色环电阻检测与定位方法

Method for Color-ring Resistor Detection and Localization in Printed Circuit Board Based on Convolutional Neural Network

色环电阻是印刷电路板(PCB)中最常用的电子元器件之一,主要依靠色环的排列顺序和颜色等视觉信息进行区分,易发生装配错误。但是色环电阻装配质量的人工检测方法效率低、误检率高,而传统的基于图像处理技术的自动检测方法鲁棒性较差,难以解决不同拍摄角度、物距及光照条件下的PCB板色环电阻检测问题。针对这一问题,该文提出一种基于卷积神经网络(CNN)的PCB板色环电阻自动检测与定位方法,首先采用编码器-解码器结构的卷积神经网络模型及带有权重的交叉熵损失函数的网络训练方法,较好地解决了复杂光照及场景下PCB板色环电阻的图像分割问题;然后采用最小面积外接矩形方法定位单个色环电阻,并通过仿射变换对色环电阻位置进行垂直校正;最后通过高斯模板匹配方法实现了色环电阻的色环定位。采用1270幅PCB图像对该文方法进行了实验和验证,并与传统的基于形态学和基于模板匹配的色环电阻检测方法进行了对比,结果表明,该文方法在召回率、准确率及重叠度等性能指标上具有明显优势,处理速度快,能满足实际应用要求。

The color-ring resistor is one of the most commonly used electronic components in Printed Circuit Board(PCB).It is featured by sequential color rings,which often brings assembling errors,however.Manual detection of color-ring resistors has low efficiency and high false detection rate.Traditional image-based automatic detection methods have difficulties in dealing with PCB images under various illuminations,imaging distance and views.To solve this problem,an automatic detection and localization method for PCB color-ring resistor is proposed based on Convolution Neural Network(CNN).Firstly,the encoder-decoder CNN model is established and trained using weighted cross-entropy loss function.With CNN,color-ring resistors are segmented from PCB images with complex illumination and scenes.Secondly,each color-ring resistor is localized using minimum area bounding rectangle,and its position is adjusted to the vertical direction by affine transformation.Finally,the localization of color rings on the resistor is achieved by Gaussian template matching.The proposed method is tested and verified by 1270 PCB images,and the result is compared with that of the traditional method(method based on geometric contour,and method based on template matching).It is shown that the proposed method has obvious advantages in performance indices,including recall rate,precision,and intersection of unions,which can meet the requirements of practical applications.