Deflectometric data segmentation based on fully convolutional neural networks

The purpose of this paper is to explore the use of Fully Convolutional Neural Networks (FCN) to perform a semantic segmentation of deflectometric recordings for quality control of reflective surfaces. The proposed method relies on a U-Net network to identify the location and boundaries of the object, and the possible defective areas present, by performing a pixel-wise classification based on local curvatures and data modulation. Experiments performed on a real industrial problem demonstrate that the combination of geometric and photometric information enables the identification of a wider variety of shape and texture imperfections, with predictions closely correlated with the visual impact of the defects. The research also highlights the relevance of the labeling process and human inspection limits, and suggestions are presented for a near-Term industrial utilization.