基于生成对抗网络的烟包缺陷检测算法研究OA

[Purpose]This paper aims to propose an efficient and accurate method for detecting cosmetic defects in cigarette packs based on the practical needs of the cigarette industry.[Methods]A two-layer Generative Adversarial Network(GAN)-based approach for cigarette pack defect detection is proposed.This method employs a two-layer architecture requiring only positive samples for training:the first layer uses grayscale histogram similarity comparison to rapidly eliminate easily detectable defective samples,reducing computational load;The second layer introduces GAN with spatial attention mechanisms,achieving refined defect identification by comparing generative loss between normal images and defective samples.Comparative experiments were conducted against One-Class Support Vector Machines(One-Class SVM)and Variational Autoencoders(VAE).[Results]Experiments demonstrated that the proposed method exhibits excellent performance in both true positive rate and false positive rate,achieving the optimal overall AUC value of 98.41%on the cigarette packaging defect dataset.