Single Image Dehazing Based on Two-Stream Convolutional Neural Network

Objective The haze weather environment leads to the deterioration of visual effect image, and it is difficult carry out work advanced vision task. Therefore, dehazing image an important step before execution Traditional algorithms achieve by improving brightness contrast or constructing artificial priors such as color attenuation dark channel priors, but unstable when dealing with complex scenes. In method based on convolutional neural network, network encoding decoding structure does not consider difference after spatial information lost in stage. order overcome these problems, this paper proposes a novel end-to-end two-stream for single dehazing. Method model composed feature stream high-level semantic stream. retains detailed extracts multi-scale structural features image. A auxiliary module designed between streams. This uses attention mechanism construct unified expression different types information, realizes gradual restoration clear network. parallel residual twicing proposed, which performs at stages improve model’s ability discriminate images. Result peak signal-to-noise ratio similarity are used quantitatively evaluate results each algorithm original reached 0.852 17.557dB Hazerd dataset, were higher than all comparison algorithms. On SOTS indicators 0.955 27.348dB, sub-optimal results. experiments real images, can also excellent effects.