Improved Multi-feature Fusion Approach Based on End-to-end Convolutional Neural Network for Dehazing

This paper presents an improved multifeature fusion approach based on end-to-end convolutional neural network (CNN) for day-time and night-time hazy scenes called IMFE-Net from the point of view of the atmospheric scattering model, which directly learns the haze-free image. The multifeatures of degraded image are composed of scene depth, hue, and atmospheric light weight, which are directly extracted from each layer of the IMFE-Net. We use a synthetic dataset and real-world hazy image training set to train the multifeature fusion network model to obtain an optimized network model. We have performed an extensive test on the real-world hazy dataset. Experimental results have proven that the model has satisfactory results. The result shows that day-time and night-time hazy scenes have a strong sense of reality processing with our method. The IMFE-Net is better than previous methods in maintaining color and detail information; it also ensures that the algorithm is real-time and can be further used for computer vision system.