Wednesday 10 September 2025
A new approach to detecting anomalies in images has been developed, which could have significant implications for a wide range of fields, including healthcare, manufacturing and security.
The problem of anomaly detection is that most current methods rely on supervised learning, where large amounts of labeled data are required to train the model. However, this can be time-consuming and expensive, especially when dealing with complex datasets. Unsupervised anomaly detection methods, on the other hand, do not require labeled data but can struggle to accurately identify anomalies.
The new approach, developed by a team of researchers, addresses these limitations by exploiting the inherent learning bias in models. This bias refers to the way that models tend to prioritize learning normal patterns over abnormal ones, even when they have limited exposure to anomalous data.
To take advantage of this bias, the researchers developed a two-stage framework. The first stage involves partitioning the dataset into multiple subsets and training separate models on each one. These models are then used to generate anomaly scores for each sample in the dataset. By aggregating these scores across all models, the team was able to identify anomalies with high accuracy.
The second stage of the framework involves using the aggregated anomaly scores to filter out a cleaned dataset with reduced contamination. This is achieved by selecting only the samples that have low aggregated anomaly scores, which are likely to be normal.
The approach has been tested on a real-world industrial anomaly detection benchmark and has shown significant improvements over current state-of-the-art methods. The results could have important implications for industries such as healthcare, where accurate anomaly detection can help identify potential health risks earlier.
One of the key advantages of this new approach is its ability to handle datasets with varying levels of contamination. This is because the models are trained on separate subsets of the data, which allows them to adapt to different patterns and distributions in each subset. Additionally, the aggregated anomaly scores provide a robust measure of an sample’s likelihood of being anomalous, even when faced with noisy or contaminated data.
The researchers believe that their approach could be widely applicable across many fields, including computer vision, natural language processing and bioinformatics. They are currently exploring ways to further improve the method, such as by incorporating additional features or using different models in each stage.
Overall, this new approach has the potential to significantly improve our ability to detect anomalies in images, with important implications for a wide range of fields.
Cite this article: “Exploiting Model Bias for Improved Anomaly Detection in Images”, The Science Archive, 2025.
Anomaly Detection, Image Processing, Machine Learning, Unsupervised Learning, Supervised Learning, Dataset Partitioning, Model Aggregation, Contamination Handling, Robustness, Computer Vision.
