Semi-Supervised Learning for Defect Segmentation with Autoencoder Auxiliary Module

In general, one may have access to a handful of labeled normal and defect datasets. Most unlabeled datasets contain normal samples because the defect samples occurred rarely. Thus, the majority of approaches for anomaly detection are formed as unsupervised problems. Most of the previous methods have typically chosen an autoencoder to extract the common characteristics of the unlabeled dataset, assumed as normal characteristics, and determine the unsuccessfully reconstructed area as the defect area in an image. However, we could waste the ground truth data if we leave them unused. In addition, a suitable choice of threshold value is needed for anomaly segmentation. In our study, we propose a semi-supervised setting to make use of both unlabeled and labeled samples and the network is trained to segment out defect regions automatically. We first train an autoencoder network to reconstruct defect-free images from an unlabeled dataset, mostly containing normal samples. Then, a difference map between the input and the reconstructed image is calculated and feeds along with the corresponding input image into the subsequent segmentation module. We share the ground truth for both kinds of input and train the network with binary cross-entropy loss. Additional difference images can also increase stability during training. Finally, we show extensive experimental results to prove that, with help from a handful of ground-truth segmentation maps, the result is improved overall by 3.83%.

Automated summary

The study proposed a semi-supervised approach that utilized both unlabeled and labeled samples to automatically segment out defect regions through network training. Experimental results showed a 3.83% overall improvement with the help of a handful of ground-truth segmentation maps.