Endoscopes are effective medical devices for diagnosing diseases of human organs.However,due to the influence of the internal cavity environment of the human body,the resolution of endoscope images is generally low.Most existing deep learning-based super-resolution algorithms directly use bicubic interpolation downsampling to obtain low-resolution(LR)images from high-quality images for paired training.However,these methods will lead to texture details loss and are not suitable for medical images.To solve this problem,this paper proposes a novel de-gradation framework for medical endoscopic images.First,diverse realistic blur kernels and noise patterns are ex-tracted from real-world low-quality endoscopic images,and then a degradation injection algorithm is proposed.The extracted real blur kernels and noise degrade the high-resolution(HR)endoscopic image into a low-resolution image.In addition,this paper proposes a high-frequency guided residual dense super-resolution network,which adopts a frequency separation strategy based on dual-frequency information interaction.And a multi-level fusion mechanism is designed to embed the extracted multi-level high-frequency information into the multi-layer features of the resid-ual dense module layer by layer.This helps recover the high-frequency details and low-frequency content of the en-doscopic image.Extensive experiments on synthetic and real-world datasets show that our method outperforms the contrastive methods with better subjective and objective quality evaluations.
维普
万方数据