Research On Image Dehazing Via Generative Adversarial Learning

Image Dehazing is one of the hot research directions in the field of computer vision.With the rapid development of Artificial Intelligence(AI)technology,computer vision related technologies such as security monitoring and driverless vehicle driving are widely used in industrial scenes.These technology are highy depend on image information.Therefore,it is an important visual task to perform image dehazing on degraded images acquired by sensing equipment in hazy weather.Image dehazing aims to restore image details,improve image contrast,etc.,which can improve the accuracy of subsequent tasks.There are still many problems with the existing dehazing methods.For example,many current dehazing methods rely on the atmospheric scattering model to obtain clear images,but there are inevitable errors in the atmospheric scattering model itself and its solution process.In addition,the current algorithm is guided by objective indicators,visual quality is ignored,and unevenly distributed haze cannot be handled well,so the restoration of image visual quality needs to be improved.Moreover,the current method relies on fully labeled pairs of datasets,the method generalization ability needs to be improved.In view of the above problems,in this thesis,we mainly studies three aspects:generative adversarial learning image dehazing,adaptive distillation attention mechanism dehazing,weakly-supervised dehazing,using new machine learning methods to solve image haze removal problems,and improving image restoration performance,reducing dependence on the datasets.The main contributions of this thesis include the following three aspects:1.An enhanced Pix2Pix dehazing method based on generative adversarial learning is proposed.Most existing image dehazing methods rely on atmospheric scattering models.The modeling process simplifies the actual conditions.Solving the model process requires estimation of intermediate parameters,which brings inevitable errors.In this thesis,an embedded generation adversarial model training strategy are proposed.So that the image dehazing task can be processed as an image style transfer problem,but no longer depends on solving the atmospheric scattering model.This model can not only effectively improve the dehazing performance,but also make the details of its results more vivid and natural,in line with the habit of human eye observation.The experimental results in various aspects show that the method proposed in this thesis can be more effective in image haze removal with high robustness,strong generalization performance and vivid and real dehazing results.In the qualitative and quantitative comparison,the performance is outstanding,which is superior to the state-of-the-art methods.2.A dehazing method based on adaptive distillation attention mechanism is proposed.The distribution of haze in natural scenes is often uneven,and the current algorithm lacks discussion of the distribution of haze,which results in poor quality of the haze removal results.In this thesis,the adaptive distillation attention mechanism and and the adaptive distillation attention module based on it are proposed to distincte between dense haze and mist regions adaptively,and distillate them apart,which means more learning and processing in dense haze areas,less in mist areas.The method not only makes the results more uniform and natural,but also can be transferred to other models to improve the dehazing performance of other models.A large number of experiments show that the dehazing method based on the adaptive distillation weight modul is proposed in this thesis has superior performance in uneven hazy and dense haze scenes.The method is applicable to multiple models.3.A weakly-supervised dehazing method guided by haze concentration labels is proposed.The vast majority of current deep learning haze methods rely on fully labeled pairs of datasets.But paired data collection is difficult and costly.To decouple the constraint relationship between the paired datasets and dehazing model,in this thesis a weakly-supervised dehazing method based on generative adversarial learning is proposed.Based on the style transfer idea in the first point,the generatve adversarial training is extended to the weakly-supervised training mode.So that using unpaired data can also obtain a dehazing model with stable effects.A large number of experiments show that the weakly-supervised method proposed in this thesis can dehaze stably and effectively,which is superior to the current unsupervised/weakly-supervised popular methods,and adds a theoretical basis for the blank weakly-supervised/unsupervised image dehazing.