GAN-based statistical modeling with adaptive schemes for surface defect inspection of IC metal packages

Metal packaging is an alternative technology to guarantee the environmental resistance and the performance reliability of ICs. Surface defect inspection of IC metal packages is an indispensable process during manufacturing. Here, a statistical modeling framework is proposed based on a GAN for surface defect inspection of IC metal packages, which involves several adaptive schemes. To the best of our knowledge, we first introduce the GAN to establish a machine vision based method for surface defect inspection of IC metal packages. IC metal package images are automatically acquired by an AOI system and employed for inspection via the proposed framework. To tackle the problem of imbalanced data in real industries, the framework only utilizes qualified samples to train the GAN template generator, which can characterize the intrinsic pattern of qualified samples. Then, a weight mask scheme is proposed to suppress the interference pixels in the difference image corresponding to qualified samples. Next, an adaptive thresholding scheme is proposed to adaptively determine an appropriate threshold for each inspected sample. Finally, an image patch-based defect evaluation scheme is designed to local-to-global evaluate the surface qualities of IC metal packages. Comparison experiments indicate that the proposed framework achieves better inspection performance in terms of 3.16% error rate and 0.89% mission rate at a reasonable inspection time of 119.86 ms per sample, which is superior to some existing deep learning based inspection methods for surface defect inspection of IC metal packages.