Low-Illumination Image Enhancement for Foreign Object Detection in Confined Spaces

Low-illumination image enhancement (LIIE) in confined spaces is a challenge faced in industrial inspection and diagnostics areas. Most existing methods cannot handle extremely dark images captured in confined spaces. This article reports a residual deep curve estimation network (ResZero) to enhance such dark images for foreign object detection (FOD). To reduce information loss when enhancing extremely dark images, residuals are added to the backbone of the proposed ResZero to pass forward features in the middle layers. To solve noise and color cast problems in the enhancement process, a new loss function is defined by considering the feature-preserving and RGB channel averaging simultaneously. Moreover, the dynamic range of the public datasets has been expanded by adding extremely dark images captured in confined spaces. Experiments with various extremely dark scenes demonstrated that the enhanced image's quality average score (BRISQUE) of the proposed ResZero is 22.25 better than all other methods(0.22 higher than the second place). In FOD experiments, the proposed method is with the highest precision.