Research On CT Image Detection Of Printed Circuit Board Element Based On Deep Learning Method

Printed Circuit Board connects the different parts of electronic components and has been an important and indispensable part of various electronic equipments. However, there are many problems in the process of producing or using PCB, such as the breakage of pad and broken circuit. These problems often cause PCB to be invalidation. So a fast and accurate method of detecting and locating the disfigurement is of greatly significance for the normally working of PCB. The technology of cone-beam CT can image PCB in three-dimensional space and gains inner structure and high resolution three-dimensional image of PCB. It provides a new fast and noncontact method to detect the disfigurement of PCB. However, conventional image detection methods mostly use pixel-level feature. This kind of feature has no semantic structure and weak robustness. As a feature extraction technique raised in recent years, deep learning has condensable structure, strong ability of feature learning and can extract semantic feature of the object to be detected. It provides a powerful tool for the detection of via, track and other elements in cone-beam CT image of PCB.This subject focuses on the demand of nondestructive detection of PCB. The research is about the technique of image detection based on deep learning and aims for solving the auto-detection of via and track in PCB. The research mainly contains the technique of feature extraction using deep learning, the method of via and track detection based on deep learning and its software implement. The main work that has been done is as follows:(1) The history, appearance and development of deep learning and shallow model are reviewed at first. Then the paper makes a summary of the advantages of deep model over shallow model. The collection of data, local extremum and gradients diffusion are the main difficulties when using traditional method to train deep net. The methods for constructing and training deep belief net are introduced emphatically. The characteristic of deep belief net is also presented in the paper. According to whether having coder and decoder or not, the current deep models are classified into three types including generative model, discriminative model and decoding model. The improvements over these three types and their advantages are also introduced briefly.(2) The conventional deep belief net has a large number of parameters. In order to relieve the influence of this problem, a kind of deep model based on prior label training is presented. A method, that using prior label to train deep model, is also proposed in the paper. The generative model is used to extract the statistical features from the training data. Then the dimension of the features is reduced by principal component analysis. The training data are re-labeled by the reduced features. Then, the re-labeled data are used to train a conventional neural network. The number of parameters in this neural net can be reduced significantly if the number of neuron in hidden layer is limited to a certain degree. The proposed method is tested on four standard dataset for machine learning. The experimental results show that the number of parameters in deep belief could be decreased to about 40% of its original scale without any losses of model generalization. Aiming at the problem of why unsupervised training help the supervised learning, a point of view is put forward to explain the mechanism of unsupervised training based on the experimental results. The view points out that the outputs of unsupervised training provide more prior information. The prior labels have a better representation of the data, which could provide a fine cost function for the model. So unsupervised training can be regarded as a kind of regularization in cost function.(3) Aiming at the problem of low contrast, large noise and artifact in CT image of PCB, a method of via and track image detection is proposed based on deep model and the image detection software is implemented. The deep model is constructed with prior label training. After trained on the image samples, the model is able to classify via and twelve types of track from background image. So the deep model could detect via and track simultaneously. The output of the model can be directly used for via detection. While for track detection, sliding window is moved continuously according to the shape of track until reach the end of the track. The experimental results show that the detection method based on deep learning is robust to CT image with low contrast and large noise and has a capacity to overcome the influence of artifact. The proposed model has better accuracy and efficiency than Hough transform.