Memory-Efficient and GPU-Oriented Visual Anomaly Detection With Incremental Dimension Reduction

Teng-Yok Lee, Yusuke Nagai, Akira Minezawa; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 2908-2916

Recent studies show that the image features from pre-trained convolution neural network (CNN) can be used for anomaly detection, even without fine-tuning. A common type of methods divides the image space into patches, and estimates the distribution of CNN-based features per patch of all training data. While this types of methods can achieve high accuracies, the high dimensionality of CNN features causes overhead to both computing and storage. In this paper, we present an incremental algorithm to reduce the dimensionality of CNN features during the training. As our algorithm ultimately computes the Truncated PCA of the features, it only maintains the truncated singular values and vectors during the training. Besides, to efficiently update the truncated singular values/vectors of all patches, we further optimize the algorithm in order to fully utilize GPUs for parallel execution. We show that with our approach, we can achieve high accuracies on the texture classes of MVTec AD with small memory footprint and extreme high speed (around 200FPS) on a single GPU.