Study Of Key Technologies On Image Features Based Machine Vision Inspection For FPC

Because FPC(Flexible Printed Circuit board) has the advantage of being thin, light and flexible, it is widely used in electronic products, such as mobile phones, notebook computers, etc. In recent years there has been an FPC industrial transfer from the world to China, so in china, FPC industrial production is growing rapidly and developing in directions of miniaturezation and high integration. Traditional FPC inspection techniques based on human vision are difficult to adapt the development of modern FPC producing because of their low stability and inefficiency. It has been the trend to apply machine vision instead of human vision to inspect FPC. This paper focuses on target location, region segmentation and defect inspection problems in FPC machine vision inspection system. Several algorithms based on image features are proposed to improve inspection process. The main contents are summarized as follows.The first stage of FPC detection is to locate a target region. To fix the detection region quickly and accurately is significant to improve inspection efficiency and accuracy. This paper proposes an image registration algorithm which is based on improved SURF(Speed Up Robust Feature). This method overcomes the inefficiency of traditional template matching technology by optimizing feature descriptors and mapping strategy, can register images with large offsets and is robust to image noise, local shelters and light changes. Experiment results show that location accuracy is highly improved.Region segmentation will help to judge the integrity of FPC area and different inspection strategies can be adopted in particular areas. The segmentation algorithms based on color or texture threshold do not take the spatial distribution of pixels and features into account, so segmentation errors are prone to occur along edges and in blurred areas. This paper adopts conditional random field to model correlation information of image context, builds feature descriptors with color and texture components, in order to solve the misjudgments in blurred regions. A multinomial logistic regression function is applied to calculate potential of different classifications in one time, so a multi-class classification can be achieve in one operation instead of being divided into several two-class classification problems. It will help to reduce complexity of the model and enhance its discrimination and stability.Classic machine vision algorithms inspect defects by pixel feature comparison. This method requires special parameter configurations for different kinds of FPC, while the configuration solutions are difficult to achieve. In order to inspect FPC commonly, this paper adopts the principle of vision saliency to search defects on FPC solder pads in the way of calculating the saliency map of images. The color contrast descriptor is based on CIEL*a*b* color space and weighted with Gaussian distance. The texture contrast descriptor is based on Gabor filters. Both contrasts are computed in scale space to gain color and texture saliency maps. The color saliency maps are combined by a maximum function and the texture saliency maps are combined by a linear function. Salient regions obtained by thresholding the saliency maps using maximum entropy. The union of color and texture salient regions can indicate the location of defects. This method does not require particular color and texture parameters and is adapt to various FPCs with different characteristic.An FPC surface inspection system is developed based on the works mentioned above. The system hardware components are listed, the selection strategy is introduced, and the command edition module and defect inspection module are presented. In addition, some details of systems process are amplified. Finally, this chapter details the process of FPC inspection including target location, region segmentation and defect inspection on solder pads. Statistical data from FPC production lines which have applied this system show that machine vision achieves more efficiency and higher accuracy than human vision.