Defect contrast enhancement imaging for rail surface inspection with neural network and structured light

Optical inspection for rail surfaces usually suffers from low defect contrast and measurement errors due to the significant variation in surface reflectivity or gloss and harsh environment. This paper regards the measurement error region such as overexposure/underexposure and data loss as a special cause for low contrast, namely, the contrast value is zero. To improve the defect contrast, this paper proposes a structured light (SL) illumination method and an unsupervised measurement-informed neural network (MeaNet). Quantitative relationships between the proposed SL parameters (e.g. gradient, period) and typical defect characteristics (e.g. crack width, crack orientation, rail surface reflectivity) are investigated. MeaNet additively fuses new information from input images and embeds defect a priori information, achieving two to three times of contrast improvement for cracks than state-of-the-art methods. Meanwhile, MeaNet generates a defect image showing only defect texture for an extremely high defect contrast. This method can be easily extended to other surface inspection applications.