Research On Printing Defect Detection Under Complex Conditions

With the improvement of living standards,people have higher requirements for the appearance quality of printed matter.The traditional manual detection is not only time-consuming and laborious,but also difficult to unify the detection standard,and there are problems of false detection and missing detection.Machine vision printing inspection can well solve the above problems,but it still faces many problems brought by complex conditions:(1)in production,the conveyor belt jitter causes the product to be inspected and the template position is inconsistent,so image calibration is needed,and the performance of image calibration will affect the detection results.The detection method based on unsupervised learning can solve this problem,but it can’t det ect small defects accurately.(2)Uneven illumination results in the difference of g ray value between the image to be detected and the template image,which leads to the detection error.(3)When the brightness and angle of the bronzing material are not ideal,the threshold segmentation results are different,the modeling is incomplete,and the regional miss detection occurs.In view of the above problems,this paper puts forward the following solutions :1.A defect detection method based on general adverse networks(GAN)is proposed.Firstly,an up-sampling module is added to the method to fuse the two sampling results to reduce the loss in the process of up sampling.Secondly,a self-attention mechanism is proposed,which takes the threshold segmented image as a branch to participate in the calculation of self-attention mechanism.It can improve the attention weight of the target area,obtain better detail processing,and help to obtain more complex structure and more accurate deta ils of the reconstructed image.Finally,the absolute value of the difference between the reconstructed image and the original image is taken,and then the noise distribution is learned,the noise is removed,and the interference outside the defect is eliminated,so as to obtain more accurate defect detection results.This method solves the problem of low precision of generating countermeasure network to detect subtle defects.2.A printing defect detection method based on multi-scale fusion is proposed.Firstly,the multi-scale network is trained.The convolution kernel of different sizes in the network can extract the shallow local image features of different scales,which helps to extract the texture features and weakens the influence of illumination on the detection.Secondly,the multi residual self-attention mechanism is added,which can learn the high-frequency residual features of the image,optimize the generation of texture and edge features in the image,and contribute to defect detection.The experimental results show that,compared with the existing methods,it can enhance the robustness of detection,and has certain advantages in the case of uneven illumination.3.A printing defect detection method based on stacked convolution network is proposed.First of all,by continuously placing multiple modules end-to-end together,expanding a single network,repeating bottom-up,top-down inference and prediction.Repeated bi-directional prediction is very important for the final performance of the network.The final network architecture has made progress in accuracy.Considering the amount of computation,a lightweight generator network with fewer layers is proposed.On the premise of ensuring the accuracy,the detection time is saved.Secondly,in the decoder stage,we use the channel weight block(CWB)to integrate the shallow spatial features and deep semanti c features,and gradually restore the predicted spatial significance value,which is helpful for defect detection in the case of reflective gildin g materials.Experimental results show that this method has good performance in defect detection of stamping materials,and can improve the detection speed compared with the previous methods.Experiments show that this method can solve the problems caused by complex conditions,and carry out accurate and effective defect detection.It can reduce the noise interference in the detection process,improve the detection accuracy based on unsupervised generation countermeasure network;improve the low accuracy in the case of uneven illumination;and put forward an effective method to solve the problem of regional missing detection due to incomplete modeling of stamping materials.