Improving Unsupervised Defect Segmentation by Applying Structural Similarity to Autoencoders
Convolutional autoencoders have emerged as popular models for unsupervised defect segmentation on image data. Most commonly, this task is performed by thresholding a pixel-wise reconstruction error based on an ℓ^p distance. However, this procedure generally leads to high novelty scores whenever the reconstruction encompasses slight localization inaccuracies around edges. We show that this problem prevents these approaches from being applied to complex real-world scenarios and that it cannot be easily avoided by employing more elaborate architectures. Instead, we propose to use a perceptual loss function based on structural similarity. Our approach achieves state-of-the-art performance on a real-world dataset of nanofibrous materials, while being trained end-to-end without requiring additional priors such as pretrained networks or handcrafted features.
READ FULL TEXT