In addition to the above, who can help us finding a company, which works with AI based visual inspection?
My work involves using a proprietary volumetric capture system, one that is image based, the post-processing leveraging AI to automate anomaly detection. While our capture tech is agnostic to subject matter, our focus has been to bring inspections of infrastructure, such as substations and bridges, into the modern era of VDI (virtual desktop infrastructure). Image-based ML (machine learning) involves a manual labeling process, someone who knows what what they're looking at, then annotates with various graphics tools the objects or features of interest, in the case of bridges, early detection of hairline cracks in concrete is a game of find Waldo and begs to be automated, both the feature extraction and classification, e.g. good, poor, okay. In your case, scratches, irregularities in paint quality, and dents, these three features vary in a way that points to different approaches to feature extraction, some more straight forward than others.
For dents you don't need AI, unless there's such a volume that motivates deeper understanding how these relate, in which case an engineer would be tasked with classification to then automate that. As is, a precise scan of a part, the 3d model nested inside the ideal ( not dented) or even the CAD model used to make the part, assigning different colors to one and the other allows the finest deviation to jump out.
For scratches, we use various tools to "separate the wheat from the chaff", modify the image in ways to allow subtle features (cracks in concrete down to .01mm, could be fine scratches) to be more discernible to the human eye for annotating. While ML typically requires huge datasets to train a model, one advantage of our capture system is providing uniform lighting, sensor, lens, etc., hence enabling automation using far smaller datasets.
As for detecting irregularities in paint, that's far more uphill for both a human and an AI to take on, but I submit here's yet another advantage to our approach to capture, one involving specialized lighting. Paint has color and it has a specular component, somewhere between glossy and highly matte. Measuring albedo color and albedo "roughness" is best done in component form, our scanning methodology uses polarized lighting to record separately these aspects of texture to make possible any true understanding of material reflectance properties. If an engineer reviewing co-polarized and cross-polarized datasets can more readily discern paint irregularities, if not maybe discover otherwise latent defects, then the same pattern recognition operating within the engineer's learning and assessment can be leveraged to scale up automating with ML.
Perhaps, we can help.
