Efficient anomaly detection with feature maps and clustering

Abstract Anomaly detection is critical for identifying various malfunctions or irregularities in image data. We present a highly efficient approach in terms of memory space, energy consumption and runtime using a clustering method to improve the performance of anomaly detection models by identifying anomalous and non-anomalous image features. Our model is based on pre-trained deep neural network feature maps and a clustering algorithm that is trained on the MVTec anomaly detection dataset, a benchmark dataset for anomaly detection methods with a focus on industrial inspections. The results show that the model is also capable of dealing with images of the gastrointestinal Kvasir-Capsule dataset taken via wireless capsule endoscopy. The model’s binary classification results outperform comparable anomaly detection methods on the Kvasir-Capsule dataset obtained with deep learning with a precision of 81.46 %, a recall of 76.01 % and a F1 score of 78.64 %. At the same time our model is highly efficient by using only approximately 4 million parameters during inference.