A Model Based on Universal Filters for Image Color Correction

Improving image quality is becoming an increasingly popular task, especially when working with mobile devices. One common approach to image enhancement is the use of convolutional neural networks. However, to achieve good results, such networks must be large enough, otherwise there is a risk of unwanted artifacts. In addition, large convolutional neural networks require significant computational resources, which can be a problem for mobile devices. Another disadvantage of such models is their low interpretability. Another approach to improving image quality is to use predefined filters combined with predictions in which cases they should be applied. We present a solution that follows this concept and outperforms both existing convolutional neural network-based models and filter-based methods in the task of image enhancement. Our model is easily adapted for mobile devices, since it contains only 47 000 parameters. On the RANDOM250 (MIT Adobe FiveK) dataset our method achieved a best structural similarity index score of 0.919 among small models and was three times faster than existing solutions.