Recently, anomaly detection methods that use pre-trained CNNs to extract image embeddings have attracted much attention. The objective of this research is to increase processing speed while maintaining accuracy in embedding-based anomaly detection methods. Previous methods extract features by concatenating the intermediate outputs of pre-trained CNNs. However, these features utilize the output of the deeper layers of the CNN, and the dimensionality of the features grows in proportion to the depth of the layers. This is a reason for the increased processing time. The proposed method speeds up the anomaly inspection process by efficiently reducing the dimensionality of image embeddings using the subspace generated by PCA. In evaluation experiments using the MVTec AD dataset, the proposed method achieved approximately 3.5 times faster processing speed (61.4 fps) than the previous methods PaDiM (6.9 fps) and PatchCore (17.4 fps), while maintaining a high anomaly detection accuracy of 99.0% on average AUROC. This result is more than twice faster than the typical video rate of 30 fps, achieving a near practical level of both inspection accuracy and inspection speed.