Enhanced Loss Function for the Design of Image Transform Models

Deep convolutional neural networks have triggered important progress in image-to-image mapping. The enhanced learning capabilities allow them to understand the input visual layout and perform local transformations customized to image content. This paper discusses an approach based on Generative Adversarial Convolutional Neural Networks (GAN) to help augment training datasets with generated images. The proposed design uses a loss function that integrates terms based on image quality metrics to validate the production of realistic images. This allows the configuration of neural parameters according to the visual perception of the generated images, without imposing a strict pixel-wise similarity with targets. This technique could be helpful for real applications in which supervised training should be done for pairs of input-output images that do not illustrate the same scene. The performance of the proposed approach is illustrated by a transformation of outdoor scenes from summer to winter. The experiments performed for different GAN schemes and quality metrics show improved quality of the generated images when the loss function includes loss terms related to visual perception.