Defogging method based on double-branch residual feature fusion

Abstract

The invention discloses a double-branch residual feature fusion-based defogging method, which can directly learn a mapping relationship between a foggy image and a fogless image, extract non-uniform fog residual features and realize defogging. Wherein the context space domain attention structure and the channel attention self-encoding structure respectively pay attention to the pixel space and the channel space of the feature map, feature information extracted by two branches is adaptively fused, a final fog layer residual image is output, and a discrimination network is introduced to enable a defogged image to be closer to the real impression. Experiments show that the method can provide a defogging effect superior to that of an existing method, the processing process consumes a long time, and the efficiency of outputting the haze layer by the network is subsequently optimized, so that the method can be applied to engineering practices such as automatic driving.

Description

technical field [0001] The invention belongs to the technical field of image restoration of computer vision, in particular to a dehazing method based on double-branch residual feature fusion. Background technique [0002] The images captured in bad weather will be disturbed by impurities such as haze, which will reduce the image contrast and detail visibility, and affect the subsequent computer vision algorithms. Therefore, it is of great practical significance to restore the details of haze images. Traditional image dehazing methods are mainly divided into the following two categories: [0003] 1) Based on image enhancement: ignore the theoretical reasons of image degradation, and enhance the image features to improve the color saturation and contrast of the image to achieve the purpose of dehazing. Such as wavelet transform, Retinex algorithm and histogram equalization. [0004] 2) Based on image restoration: The current dehazing methods based on physical models are as f...

AI Technical Summary

Problems solved by technology

[0004] 2) Image-based restoration: The current defogging method based on the physical model is as follows: through the local statistical irrelevance between the surface chromaticity and the medium propagation, the medium propagation image is established to achieve defogging. The fog effect is poor; the method based on the dark channel prior obtains the atmospheric light value first, and optimizes the transmittance map through guided filtering to perform image defogging. Value filtering estimates medium transmittance to achieve defogging. The method is simple but the restoration image is prone to halo effect

The above dehazing methods have certain scene limitations due to the limitations of prior content or artificial constraints

[0005] The current dehazing method based on deep learning is as follows: the medium transmission estimation network based on convolutional neural network, which realizes end-to-end dehazing by estimating the mapping relationship between the blurred image and its medium transmission; without estimating the transmission rate and atmospheric light intensity, by Lightweight convolutional network directly generates dehazed images; Residual dehazing network based on flattened convolutional dilated groups, which solves the problem of image artifacts caused by dilated convolutional blocks and deconvolution; End-to-end feature fusion attention network To directly restore the fog-free image, the defogging effect on the synthetic fog image is excellent, but the defogging effect on the real fog image is poor; the fast deep multi-region superposition defogging network has a high defogging efficiency on the real foggy image , but the defogging performance is relatively ordinary

Because the fog data set of some algorithms is a synthetic fog image, the synthesis method is based on the physical model of atmospheric scattering, and there is a significant difference in feature distribution from the real fog image in nature. Uneven fog and dense fog images have poor defogging effect, and residual fog is easy to remain in some areas

Images