The deep learning system can detect defects on smartphone housing (right image).
Deep learning also has several uses in medical device manufacturing. It can find defects, such as scratches on femoral knee implants, and inspect the package seals on Class 3 devices. Deep learning vision also ensures all components are present in packages during assembly verification, such as parts in a surgical kit. In addition to defect detection, deep learning can often classify the type of defect, enabling closed-loop process control.
When training a deep learning system, it is important to create a data set of sample images to build and train the model, starting with 30 to 50 images per defect and the same amount per good part. New images can then be added to reflect false reject and accept cases. By defining a full range of part, material and defect types, manufacturers can emphasize variability in the training set. It is also recommended to have two human experts grade images independently for validation and to confirm consensus between their judgment. It typically takes one week per defect to train the model.
The concept of garbage in, garbage out is critical when choosing the best images to train the system. It is ideal to collect image data sets of both good and bad parts under the expected lighting and optics conditions. Capturing high contrast images of difficult surfaces — such as glass and specular textured colored materials — requires custom lighting techniques, advanced imaging and precise part manipulation.