Visual Attention Mechanism Inspired Research On PCB CT Image Wire Detection

Wires in PCB(Printed Circuit Board)play a main role in the most electrical connections of circuit,hence,wire detection is the crucial process of PCB detection,and has an immediate influence on the accuracy of automatically locating and estimating faults.Recently,CBCT(Cone Beam Computed Tomography)imaging technique reputed for non-destruction,high efficiency,and high resolution provides a means of automatic PCB detection based on image,and gives rise to extensive attention.So,the research on wire detection in PCB CT image has important and practical values.However,the interference from metal artifacts and scattering noises result in the decrease in the quality of PCB CT image,such as intensity inhomogeneity that destroys compact boundary structure of wires.Because of complex electrical connections of circuit,wires have the characteristics of plenty of twists and turns and dense local distribution.Massive vias,pads and cover coppers appear together with wires and lead to mazy detecting surroundings.Wire detection belonging to line type of object detection research has been a difficult job in computer vision.Aimed at these problems,in order to improve the accuracy of location and recognition of wire detection,inspired by the human visual attention mechanism,an in-depth study has been performed based on the characteristics of low-level and high-level features,and main innovative harvests are as follows:1.This study proposes a wire detection method based on edge-preserved graph-based segmentation(EPGS)to solve the problem of severe intensity inhomogeneity in PCB images.Considering that superpixel segmentation methods employ low-level features to obtain preferable adherence to boundaries and benefit accurate location of wire,similar top-hat transformation of PCB CT image based on edge-preserved filter is firstly performed to increase the difference between wires and background,which improves superpixel segmentation to protect edges of wires.Then wire detection is effectively accomplished by means of recognizing superpixels based on line type of features,and intensity features are added to decrease false positives further.Results demonstrate that the proposed method achieves better wire detection and has higher edge protection.2.This study proposes a wire detection method based on transfer learning and FCN(Fully Convolutional Network)to solve the problem of local dense and various PCB element distribution in PCB images,considering that CNN(Convolutional Neural Network)can learn high-level features of PCB elements from annotated samples to improve the ability to recognise wires in the confused surroundings,a CNN classifier of PCB CT image patches is firstly trained to learn PCB element features to solve the problem of minor semantic segmentation dataset.Then end-to-end fine tuning FCN by transfer learning on minor dataset and wires can be detected by extracting wire regions from semantic segmentation results.Results demonstrate that the proposed method achieves better wire detection and has higher improvement in recognition accuracy compared with traditional methods without learning.3.This study proposes a wire detection method based on the bottom-up and top-down manner of visual attention mechanism in order to obtain both location accuracy and recognition accuracy.In terms of visual stimulus,people firstly employ visual cortices to process visual information and obtain global recognition and understanding in a bottom-up manner,then local and detailed analysis of visual stimulus is performed based on semantic understanding in a top-down manner.Additional considering the advantages of superpixel segmentation that employs low-level features to improve location accuracy and FCN employs high-level features to improve recognition accuracy,global semantic understanding is firstly obtained by FCN in a bottom-up manner,and semantic information from FCN is introduced into image in a top-down manner,and teaches EPGS method to produce better superpixels.Then employ local consistency in superpixels to constrain semantic probabilistic maps of FCN,and improve semantic segmentation results.Finally wire probabilistic energy item is constructed and introduced into graph cut energy function to improve detecting quality further.Results demonstrate that the proposed method based on the combination of bottom-up and top-down manner and fusion of low-level and high-level features,better wire detection is achieved in terms of both location accuracy and recognition accuracy.