Evaluation of Fine-Grained Anomaly Detection Methods on a Novel Battery Surface Defect Dataset

Industrial anomaly detection frequently faces challenges, including an imbalanced distribution of positive and negative samples, complex backgrounds in product images, and the recognition of small targets. Although anomaly detection (AD) has achieved notable improvements, inadequate learning of localized reconstruction features impedes further enhancements in AD performance. In this paper, we propose the Fine-Grained Anomaly Detection Methods (FGADM). This method utilizes a Artificial Anomaly Data Generation Module to produce a significant number of augmented abnormal samples. These samples are combined with normal samples to train a Fine-Grained Anomaly Detection Network. The abnormal score maps generated by the reconstruction network during the coarse stage inform the learning process of the reconstruction network in the fine stage. Ultimately, a discriminative subnet learns the distance metrics between positive and negative samples, leading to the final anomaly scores. Furthermore, to expand the data available for anomaly detection tasks in the industrial sector, we introduce a new dataset based on optical images of batteries, named the Novel Battery Surface Defect Dataset (BSD Dataset). This dataset comprises 7,898 images, each with a resolution of 330 × 260 pixels, primarily featuring defects such as Dust, Stain, and Scratch. The results of the experiments reveal that our method exceeds existing leading-edge methods on the BSD Dataset, showcasing remarkable effectiveness.