Underwater Image Enhancement for Automotive Wading

To improve the driver safety for automotive wading, the underwater image enhancement is designed to help drivers to deal with fuzzy images shown in highly scattered underwater environments and detect hidden obstacles. In addition to the traditional methods, some neural network approaches which may perform better are suggested. The purpose of this paper is to integrate traditional methods and deep learning methods to enhance underwater images. The former Deep Underwater Image Enhancement Network (DUIENET) employs a gated fusion network architecture to learn three confidence maps which will be used to combine the three input images into an enhanced result, the white balance is replaced by Multi-Scale Retinex in the modified approach. Then, some objective evaluation methods are also adopted, mean square error method and peak signal to noise ratio method, to compare the performance of the different methods. Finally, we chose some test images from test data randomly, the results show that the developed approach is better than the previous methods, the clarity of the pictures have been indeed improved.