Enhancement of Weakly Illuminated Images Using CNN and Retinex Theory

摘要

Images captured in low light environments are more susceptible to loss of information in dark regions, making details of the scene not noticeable to humans. These dark images can also make difficult the use of automatic computer vision algorithms in applications as segmentation, object detection and recognition, or tracking. This paper proposes a Convolutional Neural Network (CNN) based method for the enhancement of weakly illuminated images. The network architecture estimates the illumination of the scene which is further used to enhance the images using the Retinex model. The experiments conducted in datasets with synthetic and natural images proved that our method surpassed other state of the art approaches (quantitatively and qualitatively), creating images with less color distortion.