Research On Defect Inspection Of IC Metal Package Based On Generative Adversarial Network

In order to ensure the environmental resistance and performance reliability of integrated circuit（IC）,IC packaging is an indispensable process in the IC manufacturing.However,since unstable factors commonly emerge in the manufacturing process,contaminations and even scratches inevitably appear on the surface of the IC package.These defects will erode and affect the packaging surface to a certain extent,which further affects the stability and reliability of the IC chip.Therefore,surface defect detection of IC packages is valuable and indispensable for IC manufacturing.This thesis proposes two surface inspection methods for IC metal packages,which deal with the integration of deep learning and statistical modeling,and the fusion of multi-scale templates.They can not only suppress the inherent interferences caused by template reconstruction,but also construct templates adaptively with qualified samples.They can solve the problems of the diversity of defect appearances for IC metal packages,the difficulty of collecting a large number of negative samples,and the unbalanced data,which is beneficial for practical industrial applications.The work mainly includes the following contents.（1）In order to suppress the inherent error interferences existed in previous inspection methods,we propose a defect inspection framework for IC metal packages,which integrates statistical modeling and generate adversarial templates with adaptive mechanisms.The framework includes three stages,such as generator training,offline statistical modeling and online real-time inspection.In the stage of offline statistical modeling,qualified samples from the training set are input into the trained generator to obtain the corresponding templates.Then,an average feature map is achieved by the difference image between each qualified sample and its template to characterize the inherent reconstruction error.Next,a corresponding weight mask map is constructed by the average feature map.In the stage of online real-time inspection,to highlight the defect pixels for each inspected image,we adaptively determine the thresholds for different defect pixels according to the mean and standard deviation of the difference image for each inspected image.Also,we design a local-to-global evaluation method for defect assessment.Experimental results show that the framework can segment the defect pixels accurately,which proves its excellent surface defect inspection performance.（2）To suppress the inherent reconstruction error existed in the data-driven generated templates,we propose a surface inspection framework with the fusion of multi-scale templates for IC metal packages.By down-sampling the image to be detected,the images with different scales achieved by downsampling the inspected image are input into the corresponding networks to achieve the corresponding templates for different scales.The networks for different scales pay attention to different aspects,for example,the network for large scale is good at extracting subtle features.Thus,multi-scale defect maps are achieved with different attentions.The reconstruction errors are suppressed by evaluating the image blocks to obtain a defect pixel map with the same scale to the original input scale.Finally,quality assessment is globally performed.Experimental results show that the proposed method improve the inspection performance to a certain extent with a slight decrease of the time efficiency.