A lot of research has been conducted on image restoration, which is a problem of restoring a distorted image to a clean image. Recently, Many tasks in the image restoration field are showing tremendous performance improvement due to the advancement of deep learning.
In particular, for real-world applications such as autopilot vision or multi-camera sys- tems, some datasets using multiple distortions have been studied in recent years. Based on the method of applying multiple distortions to the images, these datasets are broadly classified into two types: sequentially applied to the entire image, and applied differently for each divided section of the image. The model trained through the dataset created in the first method is weak in restoring a single distortion images, and the model trained in the dataset created in the second method is weak in restoring the mixed distortion images.
Therefore, we propose the Integrated Multiple Distortion Dataset by fusing these two types of methods. The proposed IMDD is a useful dataset that can be trained in sit- uations where multi-distortions are assumed by supplementing the shortcomings of the two datasets. Additionally, we propose a Spatial Distortion Information Transfer Net- work(SDITNet) consisting of a reconstruction network that restores a multi-distorted image and a recognition network that predicts the spatial distortion information of the multi-distorted image to deliver it to the reconstruction network. We find that we show that our model outperforms other models for both single-distortion and multi-distortion reconstructions, and that the transfer of spatial distortion information obtained through the recognition network helps to improve the reconstruction performance.
