Research on Long-Distance Sea Surface Image Dehazing via Neural Network

This research focuses on the problem of dehazing long-distance sea surface images. The aim of our study is to solve the problems that traditional methods cannot deal with in hazy sea and sky images, while simultaneously improving the speed and accuracy of neural network image dehazing. In terms of methodology, we first used the traditional Dark Channel Prior dehazing method to dehaze images of the sea and sky. Next, the network proposed by MB-TaylorFormer: Multi-branch Efficient Transformer Expanded by Taylor Formula for Image Dehazing was used for training and experiments. Finally, the experimental results of the two methods were compared and discussed. We created a training data set by adding a white layer and adjusting transparency to simulate fog. And we used a prepared data set to adjust the weight and base of the model. This allows the neural network to process sea surface image dehazing. Additionally, we have experimented with some of the images from this training set using the traditional dark channel method. We conducted experiments on several created dehazing datasets. Experimental results show that the MB-TaylorFormer network outperforms other schemes with a low number of parameters and computational complexity. In particular, the results are more significant when compared with traditional methods. The average PSNR improvement is approximately 47.73% and the average SSIM improvement is approximately 19.61% relative to the traditional method.