Model-driven photometric stereo for in-process inspection of non-diffuse curved surfaces

This paper presents an in-process inspection approach for quality control of non-diffuse curved surfaces based upon an evolution of conventional photometric stereo. An inverse reflectance model is proposed to reveal the non-linear reflectance behaviour of general non-diffuse surfaces from images based on a neural network. The model can directly be used to drive a high accurate photometric stereo which only requires a single RGB image with a pre-captured collocated image. This allows the technique to realize in-process inspection of moving surfaces with micro-second level capturing rate. Experimental study confirms the excellent texture recovery and defect detection capabilities in mass production.