Cellular Data Generator Based Siamese Container Marking Anomaly Detection Network

Container marking anomaly detection is a pivotal component within container production processes. Due to the diversity and variability of container markings, manual detection is inefficient and prone to errors. While mainstream deep learning-based anomaly detection methods typically focus on products with consistent appearances, container markings from different production batches often exhibit variations. To address this, we propose a Cellular Data Generator Based Siamese Container Marking Anomaly Detection Network (GS-CMAD). GS-CMAD is trained on data generated using our method during the training phase and employs a siamese network for anomaly detection, addressing issues of scarce anomaly data and detection of unknown class markings. Additionally, we propose a novel Cellular Data Generator (CDG), utilizing Cellular noise to generate data that closely resemble the distribution of real container marking data compared to methods using Perlin noise. To enable the generated data for training siamese networks, we propose a novel method, Data Generation for Siamese Network (DGSN). DGSN consists of two CDGs with identical parameters, capable of simultaneously generating normal samples and anomaly samples, which can be directly used for siamese network training. Subsequently, we propose the Siamese Container Marking Anomaly Detection Network (SADN), utilizing siamese networks to extract difference features between template and target images for anomaly detection. Additionally, due to the absence of corresponding datasets, we proposed the Container Marking Anomaly Detection and Localization Dataset (CM-ADL). Finally, through extensive experimentation, we have validated the effectiveness of SADN and CDG, thus demonstrating the advancement of GS-CMAD.