Feature Decoupled Autoencoder: Semi-Supervised Learning for Image Dehazing

Abstract

Single image dehazing plays an important role in image processing. Now, most image dehazing methods only use synthetic datasets to train models which produce some poor results on the real hazy images. To solve this problem, we propose a new network named feature decoupled autoencoder which can be trained by semi-supervised learning due to the real hazy images lacking labels. Our autoencoder includes a feature decoupled encoder and a feature fusion decoder. The clear image feature and the hazy feature are decomposed by the encoder and merged by the decoder. When the hazy feature is set to 0, the decoder will produce a dehazing image relying on the clear image feature. By training on the synthetic datasets and real datasets together, our model has good dehazing ability on the real-world hazy images. On the HazeRD and SOTS, our model's performance is comparable to the state-of-the-art algorithms.