Unsupervised Infrared Single Image Super-resolution Method Based on Double Discriminator Generative Adversarial Network

Abstract

The present invention relates to the unsupervised infrared single-image super-resolution based on double discriminator generation confrontation network, including 1. Constructing a learning framework, constructing an unmatched unsupervised learning framework, including generating confrontation network and content constraint module, generating confrontation network Including the generator and the discriminator, the learning framework extracts infrared features from the image and generates a realistic super-resolution infrared image; 2. Build a double discriminator structure, 3. Combine modules, 4. Create a data set. The invention makes the super-resolution image have the advantages of low noise, clear edge texture, and high contrast; the degraded image of the super-resolution image and the pre-interpolated and enlarged image retain and maintain low-frequency information; While reconstructing the video information, maintain the harmony and unity of the overall image style and avoid the generation of abnormal pixels; improve the structure of the discriminator, and use the pixel-level true and false discrimination matrix to enhance the reconstruction of texture detail information.

Description

technical field [0001] The invention relates to an unsupervised infrared single image super-resolution method based on double discriminator generation confrontation network, which belongs to the technical field of image processing. Background technique [0002] With the advancement of computer technology, the demand for high-quality infrared images in various fields is increasing, such as: video surveillance, medical diagnosis and remote sensing, etc. However, due to the limitations of infrared sensor technology, such as infrared optical diffraction effects and uneven sensor response, infrared images acquired by infrared imaging systems often have low resolution, high noise, and poor contrast; Visible light imaging systems are expensive, making widespread adoption difficult. Improving the hardware design, such as increasing the size of the photoreceptor and reducing the pixel size, is the most fundamental and direct way to improve the imaging quality, but it requires a long...

AI Technical Summary

Problems solved by technology

However, the CinCGAN network relies on the existing supervised super-resolution network model for tuning training. The constraints similar to iterative back-projection in the dSRVAE network may lead to blurred images. In addition, the two are researches on visible light images, and the improvement in vision and indicators must also be achieved. Benefit from the noise reduction module before the super-resolution module, and the noise reduction module is not suitable for infrared images with low resolution, poor contrast, and low signal-to-noise ratio, so the unsupervised super-resolution algorithm of infrared images that does not depend on paired data sets There is still a lot of room for research and breakthroughs

Images