Deep Learning-Enabled Low-light Image Enhancement In Maritime Video Surveillance

With the rapid development of artificial intelligence technology and computer vision technology, the maritime camera is gradually applied to monitoring the inland river bridge area. However, the monitoring images of inland river bridge area obtained under low light imaging conditions usually have the characteristics of low brightness, low contrast, and low resolution, which brings many restrictions on the supervision and management of inland river traffic. To further solve the problem of the poor imaging effect of the low light images, this paper proposes a low light image enhancement method based on deep learning. Four typical neural network structures (i.e., Attention-guided CNN, DnCNN, DenseNet, and Noise2Noise) are applied to mapping the low light image to the positive light image, and then an EN-Net (Enhancement Network) is used to balance and enhance the output results of these four networks, to get a final output image that has been splined and enhanced. EN-Net will adopt U-Net's basic structure and integrate the local residual and global residual modules to increase the diversity of features and speed up training. L1 norm is used as the loss function to improve the quality of enhanced images further. The experimental results show that the depth network proposed in this paper improves the brightness and contrast with the monitoring images of the inland river bridge area and further improves the monitoring effect of the inland river bridges area, thus providing a guarantee of water traffic safety in the bridge area to a certain extent.