Blemish Detection Algorithms for Image Sensor Improvement and Qualification During Camera Production

Ensuring the quality of image sensors is critical in modern imaging applications, yet traditional blemish detection methods are time-consuming and struggle to scale with increasing sensor resolutions. This paper proposes the Isolation Forest model for efficient and scalable anomaly detection in image sensors. Using a real-world labeled dataset of RAW images provided by Bosch, we extract basic statistical features from the top 10% of pixel intensities to differentiate between defective and non-defective sensors. Despite the severe class imbalance and limited sample size, the iForest model achieved excellent results, with F1 score of 91.3% and perfect recall of 100% on the test set. The proposed method reduces drastically the inspection time (the anomaly can be detected in a less than a second) offering a promising alternative to traditional inspection techniques in high-throughput manufacturing environments.