Zero-sample unsupervised image defogging method and system

Abstract

The invention discloses a zero-sample unsupervised image defogging method and a zero-sample unsupervised image defogging system. The zero-sample unsupervised image defogging method comprises the following steps: acquiring a foggy image; the constructed zero-sample unsupervised defogging neural network comprises an atmosphere illumination image prediction branch, a transmission image prediction branch and a fogless image prediction branch; the foggy image is subjected to atmospheric image and transmission image prediction branches to obtain an atmospheric light value image and a transmission image, and regularization loss is calculated; calculating the atmospheric light loss of the atmospheric image prediction branch through the variational auto-encoder loss and the atmospheric light value constraint loss; the foggy image is processed by a fogless image prediction branch to obtain a fogless image, and the loss of the defogged image is calculated based on prior knowledge of statistical characteristics; calculating a reconstructed fog image and calculating reconstruction loss; the total loss is the sum of reconstruction loss, defogged image loss, atmospheric light loss and regularization loss; and network parameters are updated through loop iteration, network training is completed, and a final defogging result is obtained. According to the method, the inference speed of unsupervised defogging is improved, and the defogging effect is better.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a zero-sample unsupervised image dehazing method and system. Background technique [0002] Smog refers to smoke, dust or moisture in the atmosphere. With the intensification of environmental pollution, haze weather becomes more and more frequent. The presence of haze can significantly reduce image quality, resulting in poor performance in various image processing tasks such as image recognition and detection, remote sensing, and video surveillance. Therefore, image dehazing has attracted much attention as a key technique to restore degraded images captured in severe weather. [0003] Currently, image dehazing algorithms are mainly divided into two categories: prior-based methods and learning-based methods. Prior-based methods introduce reasonable prior knowledge as a constraint, which has the effect of naturalness and high fidelity. However, the effect of such methods ...

AI Technical Summary

Problems solved by technology

[0005] In order to overcome the defects and deficiencies of the existing technology, the present invention provides a zero-sample unsupervised image defogging method and system. The present invention only uses a single foggy image for training, and introduces a variety of priors for the loss function. Optimizing the loss function and training parameters, on the one hand, solves the dependence of the defogging method on prior knowledge and large-scale paired real data sets; Single image inference time

Images