Dissolving Is Amplifying: Towards Fine-Grained Anomaly Detection

• DOI: 10.48550/arxiv.2302.14696
• Publication Date: 2023-01-01
• Authors: Jian Shi; Pengyi Zhang; Ni Zhang; Hakim Ghazzai; Yehia Massoud
• Keywords: Benchmark (surveying); Discriminative model; Feature (linguistics); Code (set theory); Anomaly (physics); Representation; Feature Learning

In this paper, we introduce \textit{DIA}, dissolving is amplifying. DIA a fine-grained anomaly detection framework for medical images. We describe two novel components in the paper. First, \textit{dissolving transformations}. Our main observation that generative diffusion models are feature-aware and applying them to images certain manner can remove or diminish discriminative features such as tumors hemorrhaging. Second, an \textit{amplifying framework} based on contrastive learning learn semantically meaningful representation of self-supervised manner. The amplifying contrasts additional pairs with without transformations applied thereby boosts feature representations. significantly improves performance around 18.40\% AUC boost against baseline method achieves overall SOTA other benchmark methods. code available at \url{https://github.com/shijianjian/DIA.git}