SNR-optimized image fusion for transparent object inspection

Automated visual inspection of transparent objects is important for many industrial fields. Especially the detection of scattering impurities inside complexly shaped transparent objects is a demanding task. Usually, so-called dark field approaches are employed in this case. However, these methods often fail due to direct reflections of the light sources, e.g., at the test object's surface which cannot be distinguished from signals of real material defects. This paper introduces an inspection approach which captures images at different illumination modalities and fuses them while optimizing the signal-to-noise ratio. Two fusion strategies are presented, which employ prior knowledge in order to obtain optimized inspection images. The signal component of the observed images is defined as the signal corresponding to visualized defects. Conversely, all light reaching the sensor due to scattering or reflections caused by the test object's geometry is regarded as noise. The signal values and noise values depend on both the pixel position and the respective illumination source. Prior knowledge about the signal and noise components allows to estimate the spatially resolved SNR for every illumination channel. The images resulting from the fusion step show scattering material defects with high contrast whereas surface reflections are nearly completely mitigated by the SNR-optimized fusion strategies. Several experiments state the performance of the presented approaches.