ABSTRACT
Single image super-resolution (SISR) refers to the reconstruction from the corresponding low-resolution (LR) image input to a high-resolution (HR) image. However, since a single low-resolution image corresponds to multiple high-resolution images, this is an ill-posed problem. In recent years, generative model-based SISR methods have outperformed conventional SISR methods in performance. However, the SISR methods based on GAN, VAE, and Flow have the problems of unstable training, low sampling quality, and expensive computational cost. These models also struggle to achieve the trifecta of diverse, high-quality, and fast sampling. In particular, denoising diffusion probabilistic models have shown impressive variety and high quality of samples, but their expensive sampling cost prevents them from being well applied in the real world. In this paper, we investigate the fundamental reason for the slow sampling speed of the SISR method based on the diffusion model lies in the Gaussian assumption used in the previous diffusion model, which is only applicable for small step sizes. We propose a new Single Image Super-Resolution with Denoising Diffusion GANS (SRDDGAN) to achieve large-step denoising, sample diversity, and training stability. Our approach combines denoising diffusion models with GANs to generate images conditionally, using a multimodal conditional GAN to model each denoising step. SRDDGAN outperforms existing diffusion model-based methods regarding PSNR and perceptual quality metrics, while the added latent variable Z solution explores the diversity of likely HR spatial domain. Notably, the SRDDGAN model infers nearly 11 times faster than diffusion-based SR3, making it a more practical solution for real-world applications.
ABSTRACT
To alleviate the social contradiction between limited medical resources and increasing medical needs, the medical image-assisted diagnosis based on deep learning has become the research focus in Wise Information Technology of med. Most of the existing medical segmentation models based on Convolution or Transformer have achieved relatively sound effects. However, the Convolution-based model with a limited receptive field cannot establish long-distance dependencies between features as the Network deepens. The Transformer-based model produces large computation overhead and cannot generalize the bias of local features and perceive the position feature of medical images, which are essential in medical image segmentation. To address those issues, we present Triple Gate MultiLayer Perceptron U-Net (TGMLP U-Net), a medical image segmentation model based on MLP, in which we design the Triple Gate MultiLayer Perceptron (TGMLP), composed of three parts. Firstly, considering encoding the position information of features, we propose the Triple MLP module based on MultiLayer Perceptron in this model. It uses linear projection to encode features from the high, wide, and channel dimensions, enabling the model to capture the long-distance dependence of features along the spatial dimension and the precise position information of features in three dimensions with less computational overhead. Then, we design the Local Priors and Global Perceptron module. The Global Perceptron divides the feature map into different partitions and conducts correlation modelling for each partition to establish the global dependency between partitions. The Local Priors uses multi-scale Convolution with high local feature extraction ability to explore further the relationship of context feature information within the structure. At last, we suggest a Gate-controlled Mechanism to effectively solves the problem that the dependence of position embeddings between Patches and within Patches in medical images cannot be well learned due to the relatively small number of samples in medical images segmentation data. Experimental results indicate that the proposed model outperforms other state-of-the-art models in most evaluation indicators, demonstrating its excellent performance in segmenting medical images.
