Anomaly Detection by Deep Learning Named 'Sense Learning'

摘要

In a visual inspection setting, quantitative and qualitative judgment is needed. To automate this particular judgment process, an algorithm that resembles human recognition ability is required. Sense Learning is our algorithm that mimics the human recognition ability using the images of non-defective products. With an autoencoder, a deep-learning-based network, the inspection AI progressively learns to reconstruct the images of non-defective products, thereby acknowledging the ideal characteristics of products. When the image of a defective product is input, the inspection AI fails to reconstruct the defective part because it has never learned the form. Based on the difference between the input image and the reconstructed image, Sense Learning detects the defective part and measures the degree of its severity.