Towards Multi-domain Single Image Dehazing via Test-time Training

Recent years have witnessed significant progress in the area of single image dehazing, thanks to the employment of deep neural networks and diverse datasets. Most of the existing methods perform well when the training and testing are conducted on a single dataset. However, they are not able to handle different types of hazy images using a dehazing model trained on a particular dataset. One possible remedy is to perform training on multiple datasets jointly. However, we observe that this training strategy tends to compromise the model performance on individual datasets. Motivated by this observation, we propose a test-time training method which leverages a helper network to assist the dehazing model in better adapting to a domain of interest. Specifically, during the test time, the helper network evaluates the quality of the dehazing results, then directs the dehazing network to improve the quality by adjusting its parameters via self-supervision. Nevertheless, the inclusion of the helper network does not automatically ensure the desired performance improvement. For this reason, a metalearning approach is employed to make the objectives of the dehazing and helper networks consistent with each other. We demonstrate the effectiveness of the proposed method by providing extensive supporting experiments.

1. Introduction

Single image dehazing is a classic but still active research topic in low-level computer vision, which aims to restore clean images from the degraded hazy counterparts. Recently, many deep learning approaches [5], [10], [14], [22], [25], [26], [31], [35], [45], [49], [5]0 have been proposed to address this problem by training a neural network to approximate the mapping from hazy images to haze-free ground truths. As more and more dehazing datasets have been released, such as RESIDE [23], O-Haze [3] and NH-Haze [2], these methods are able to demonstrate their outstanding ability in handling different haze patterns. However, one important issue is left behind for consideration, i.e., handling different types of hazy images by a single network. To be specific, current methods are usually trained on the training split of a particular dataset and tested on the corresponding testing split. For example, the test accuracy on RESIDE indoor test set [23] is obtained by validating a dehazing model trained on the RESIDE indoor training set. Such an evaluation strategy allows the neural network to focus on a specific domain but evades the important problem of learning a general model across datasets. A seemingly simple remedy is to train a single dehazing model on all available datasets jointly. Intuitively, with the increase of data, the network can benefit from considering more kinds of haze patterns, leading to boosted performance on every single dataset [1]. Figure 1:

Average psnr values of gdn [26], msbdn [11] and dw-gan [14] across four datasets. It can be observed that the dehazing methods perform better if the training and validation are conducted on a single dataset.