PCB Bare Board Defect Detection Based On Improved ORB Image Registration And Deep Learning

Today,with the rapid development of the manufacturing industry,the traditional method of detecting defects through artificial vision has gradually faded out of the industrial detection environment,and detection methods based on machine vision have emerged and are widely used in industrial environments.The production of bare PCB boards is the most basic in the process flow.For the existing PCB bare board defect detection schemes,there is no universality and the classification accuracy of multiple defects is not ideal.The improved ORB image registration is designed.PCB deep board defect detection methods with deep learning are mainly divided into three parts: prepositioning,defect location and classification recognition,to realize automatic detection of PCB bare board defects.In the previous part,after analyzing the common correction methods and combining the precise effects of the experiment,the median conversion is used for preprocessing.Then,in view of the serious distortion of histogram equalization and the lack of the enhancement ability of gamma transformation and linear transformation,a PCB bare board image enhancement algorithm based on RGB space was designed.After analyzing the color characteristics of the PCB bare board,by looking for the threshold value in the G channel and combining the human eye's sensitivity to the three RGB colors,the RGB three-channel image is processed by piecewise linear transformation.The experimental results show that the algorithm index of this paper reaches 57.13,and it is not much different from the traditional linear change in terms of structural similarity and gradient amplitude similarity index.And the visual improvement is obvious,achieving the effect of moderate brightness and clearness.In the defect location part,a hybrid method based on the reference comparison method and morphological processing is used to achieve defect location.Aiming at the problems that the image registration algorithm based on ORB can meet the speed requirements of the industrial environment,but its algorithm does not meet the adaptability of the scale,the mismatch rate is high,and the calculation time of the RANSAC method is not ideal,etc.,an improved ORB algorithm is proposed.PCB image registration method of bare board.First,a non-linear scale space is established,and then the corner detection is performed using the Shi-Tomasi algorithm after strong corner processing.The final result is a more efficient PROSAC algorithm to replace the traditional RANSAC to eliminate matching points for precise matching.Experimental results show that the average registration accuracy difference of this algorithm is 0.545,and the registration time exceeds 600 ms.Compared with the traditional ORB matching algorithm,it has better size adaptability,higher matching accuracy and registration accuracy,and has a certain degree of accuracy.Real-time performance.Subsequent defect localization,using local adaptive binarization,XOR operation operation and morphological processing methods,combined with the filtering small point steps and positioning strategy designed in this paper,to achieve the removal of false defects and the positioning of real defects.Experiments on defect positioning for 6 types of defects and different PCB bare board samples are carried out.The experimental results show that the method designed in this paper has good positioning performance and low false detection rate,which is in the range of 0%~0.2%,and has good practicality.In the defect classification and recognition part,the traditional defect classification and recognition method does not have the algorithmic versatility.In the deep learning,the shallow neural network has insufficient ability to express features and the dense network is not enough for small sample tasks but there are redundant features.TFDS data set and PCB bare board data set and conduct experiments.Experiments on the TFDS data set verified here in DenseNet.The average accuracy is improved by 0.34%.On the bare PCB data set,the network constructed in this paper was compared with the average recognition accuracy and detection speed of LBP + SVM,HOG + SVM,Cifar10 Net network and DenseNet network.The accuracy of the SE-DenseNet network constructed here reaches 98.36%,and the classification performance can be detected by other algorithms with a detection speed of 33.15 ms per image,which has good real-time performance.In order to verify the performance of the model on the identification of actual PCB bare board defects,the experiment was carried out using the defect image of the PCB bare board after the defect positioning has been carried out as the input sample of the trained network model.The experimental results show that the defect detection rate ranges from 0% to 0.4%,and the detection rate is low.Here,the network has good classification and recognition performance for the 6 types of defects on the bare PCB board.In summary,the PCB bare board defect detection algorithm based on improved ORB image registration and deep learning proposed here has certain versatility and high accuracy,and can effectively replace the detection of PCB bare board defects.