Fusion of multi-light source illuminated images for effective defect inspection on highly reflective surfaces

It is observed that a human inspector can obtain better visual observations of surface defects via changing the lighting/viewing directions from time to time. Accordingly, we first build a multi-light source illumination/acquisition system to capture images of workpieces under individual lighting directions and then propose a multi-stream CNN model to process multi-light source illuminated images for high-accuracy surface defect classification on highly reflective metal. Moreover, we present two effective techniques including individual stream deep supervision and channel attention (CA) based feature re-calibration to generate and select the most discriminative features on multi-light source illuminated images for the subsequent defect classification task. Comparative evaluation results demonstrate that our proposed method is capable of generating more accurate recognition results via the fusion of complementary features extracted on images illuminated by multi-light sources. Furthermore, our proposed light-weight CNN model can process more than 20 input frames per second on a single NVIDIA Quadro P6000 GPU (24G RAM) and is faster than a human inspector. Source codes and the newly constructed multi-light source illuminated dataset will be accessible to the public.