Abstract: Image anomaly detection and localization aim to identify and locate anomalies. Industrial image anomaly detection is of significant research importance for its value in enhancing quality control and automation in industrial production. With the rapid advancement and application of deep learning technologies, substantial progress is anticipated in this field. Traditional anomaly detection methods mainly rely on supervised learning, which requires a large number of defect samples with accurate annotations for training. However, in real-world industrial scenarios, anomalous samples are often scarce and may exhibit diverse features ranging from subtle imperfections (e.g., minor scratches) to major structural defects (e.g., missing parts or components). This poses big challenges to anomaly detection and localization. In contrast,unsupervised anomaly detection algorithms can establish a detection model without any labeled sample. Through the analysis of information and characteristics inherent to normal data, the model is able to accurately detect anomalies. In this review, we start with introducing the concept of anomalies and their wide application in industrial scenarios, and discuss the challenges in anomaly detection and localization, including data scarcity and the heterogeneity of anomaly presentations. Then, we provide an comprehensive analysis of various unsupervised learning frameworks, such as teacher-student networks,feature memory repositories, generative networks, and large-scale visual language models, along with their evolution and development in industrial anomaly detection. Moreover, we list popular datasets and evaluation metrics currently used in industrial anomaly detection as reference for researchers. Finally, we summarize existing deep learning methodologies and envisions future research directions to address current challenges and advance practical applications.