Attention-optimized deep coding and decoding defogging generative adversarial network

Abstract

Aiming at the problems that the existing defogging algorithm is difficult to process non-uniformly distributed fog and a deep convolution defogging network excessively depends on a large-scale data set, the invention provides an attention-optimized deep coding and decoding defogging generative adversarial network which adopts an encoder with a four-layer down-sampling structure. Semantic information lost due to fog is fully extracted so as to restore a clear image. Meanwhile, in a decoder network, an attention mechanism is introduced, and weights are distributed to different pixels and channels in a self-adaptive manner, so that non-uniformly distributed fog is processed. And finally, a framework of the generative adversarial network enables the model to obtain a better training effect on a small sample data set. Experimental results show that the technical scheme not only can effectively remove non-uniformly distributed fog in a real scene image, but also can recover a clear image for a real scene data set with fewer training samples, and the evaluation index is superior to that of other widely adopted comparison algorithms.

Description

technical field [0001] The technical solution belongs to the field of image processing, in particular, it is an attention-optimized deep codec defogging generative confrontation network applied to image processing. Background technique [0002] The scattering phenomenon produced by light passing through suspended particles such as fog and haze degrades the image collected by the imaging sensor, and thus loses a lot of effective information, which directly affects the performance of downstream advanced vision tasks. The purpose of image defogging is to eliminate the influence of atmospheric environment on image quality, increase the visibility of images, and provide support for downstream semantic tasks (such as image classification, object detection, etc.). [0003] Image dehazing is mainly divided into traditional dehazing and deep learning dehazing algorithms. Most traditional dehazing algorithms use prior knowledge to estimate parameters to restore the original image [1,...

AI Technical Summary

Problems solved by technology

Although the traditional algorithm has low complexity, it also has the following unavoidable defects: First, due to the superposition of errors in parameter estimation and the inaccuracy of the atmospheric scattering model, the generated dehazed image cannot completely restore the reference image

Second, the performance of the traditional algorithm is largely limited by the detection accuracy of the fog area. While removing the fog, the original low-frequency information in the fog-free area may also be removed.

[0006] First, the performance of a completely end-to-end network is overly dependent on the training results on large-scale datasets. The fog data set is difficult to obtain, which limits the performance of the model;

[0007] Second, most of the existing methods process the pixels on the picture indiscriminately, and cannot cope with the uneven distribution of fog in real scenes.

Images