Improved Training of CAE-Based Defect Detectors Using Structural Noise

Appearances of products are important to companies as they reflect the quality their manufacture customers. Nowadays, visual inspection is conducted by human inspectors. This research attempts automate this process using Convolutional AutoEncoders (CAE). Our models were trained images non-defective parts. Previous on autoencoders has reported that accuracy image regeneration can be improved adding noise training dataset, but no extensive analyse factor been done. Therefore, our method compares effects two different patterns efficiency: Gaussian and made a known structure. The test datasets comprised “defective” Over experiments, it mostly observed precision CAE sharpened when noisy data during phases. best results obtained with structural noise, defined shapes randomly corrupting data. Furthermore, able had slightly positions rotations compared ones found in dataset. However, shortcomings appeared “regular” spots (in data) partially, or totally, overlapped.