Residual Squeeze-and-Excitation Network for Battery Cell Surface Inspection

Anomaly detection remains a challenge in industrial inspection, due to difficulty in designing suitable features for classical methods and in collecting sufficient samples for deep learning based methods. We propose a novel Residual Squeeze-and-Excitation network to discover anomalies and inspect the quality of adhesives on battery cell surfaces. Owing to a compact architecture design and the utilization of an attention mechanism based module, our network generalizes well with a small amount of samples. A proper training setup further ensures our network a satisfying performance on a dataset constructed by ourselves. The network manages to accurately and robustly judge the existence of anomalies and adhesives and provide visual localization of them in an image of the battery cell surface.