AFUNet: Arbitrary Feature Upscaling for medical images super-resolution
Abstract
Deep learning has achieved remarkable success in image super-resolution, producing visually convincing high-resolution results from low-resolution inputs. However, most existing methods are designed for fixed upscaling factors, limiting their flexibility and adaptability to arbitrary magnification scales. In this work, a new model named AFUNet, a deep convolutional neural network that can learn arbitrary scales within a single model, is introduced. We make the following contributions: First, we proposed an upsample block to upscale image features more effectively, rather than only upscaling at the end of the model to produce the final output. Secondly, we use multiple feature enhancement blocks to enhance the upscaled features before generating the super-resolution images. The model is not only used as a standalone model but can also be combined with other architectures to improve the model’s performance. The experiments on two medical datasets showed that the proposed model can have competitive performance with state-of-the-art models while having up to 42% fewer parameters.