Image Dehazing Methods Combining Data Learning and Physical Priors

Abstract

The invention belongs to the field of image processing and discloses an image defogging method combining data learning and physical prior. First, the residual network is trained based on the database containing the estimated transmittance and accurate transmittance of the fog map; secondly, for a foggy image, the transmittance is estimated by the edge-limited method and the network prior is initialized with it, and for the network thereafter For each layer, the output result of the previous layer is used as its prior, and after multi-layer iterative convolution and nonlinear activation, a more accurate estimate of the transmittance is output; finally, the final transmittance is substituted into the defogging physical model to solve the defogging image. Advantages: A brand-new deep network for haze removal is established, and more accurate estimation results are obtained with less training data; physical laws are introduced into the deep learning framework to assist the deep model to describe the transmission process of the transmittance, making up for the The gap between domain knowledge and training data; better results have been obtained on real data, and thicker haze can be effectively removed.

Description

technical field [0001] The invention belongs to the field of computer vision and relates to low-level image processing, in particular to an image defogging method combining data learning and physical prior. Background technique [0002] With the increasing air pollution, the significance of improving the quality of foggy images is self-evident, and the correction and optimization of photo quality has become an obvious demand. Image dehazing is a typical representative in the intersection of image enhancement and image restoration. Early research has proposed different types of visual priors to solve this problem, the most representative of which is the dark channel-based method proposed by Dr. He Yuming in 2009. Dehaze method for estimating transmittance, but makes assumptions that may be invalid on certain images. Although the recent neural network method also shows good performance in dehazing, due to the disappearance of physical rules in the fog image, the training of t...

AI Technical Summary

Problems solved by technology

Although the recent neural network method also shows good performance in defogging, due to the disappearance of physical rules in the fog map, the training of the network requires a large amount of data, and if the fog concentration of the test image is different from that of the training data, the result is often unsatisfactory.

Images