Image dehazing with an untrained neural network
Zhiwei Li, Xinjie Xiao, Yuanhong Ren, Nannan Zhang, Hongtao Shan, Wuneng Zhou · Proceedings of the 2021 5th International Conference on Electronic Information Technology and Computer Engineering · 2021
The haze floating in the atmosphere scatters and refracts atmospheric light, which leads to degradation of image quality. Recently, many learning-based dehazing methods have been proposed, which use hazy-clean image pairs to train network models. However, it is difficult to capture the hazy-clean image pairs. To overcome this difficulty, here we propose an unsupervised dehazing model using color attenuation prior (IDUNet), which avoids collecting hazy-clean image pairs, and need not use a lot of data to train parameters before dehazing. In brief, IDUNet only needs to input a haze image and can output a haze-free image. Then, IDUNet generates a handcrafted haze image using the haze-free image and an atmospheric scattering model (ASM). The haze-free image is optimized by iteratively minimizing a specific loss function scheme that forces the handcrafted haze image to imitate the input haze image. In-depth experiments show that IDUNet can dehaze without a training model and recovers satisfactory images.