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