Research On Haze-Prior Induced Deep Dehazing Model

Image dehazing is an active research direction in the field of computer vision,receiving a growing concern in recent years.Under the influence of haze,the image quality will be decreased.These image degradations bring great difficulties to subjective appreciation and subsequent visual tasks,such as target detection,recognition,and semantic segmentation.Therefore,image dehazing tasks,which could be considered as an image restoration task,have a wide range of demands in daily life,public safety and military applications.Traditional image dehazing methods solve the problem via handcrafted priors.However,these methods lack robustness since handcrafted priors are inevitably violated in practice.Some methods use deep neural networks to directly learn the mapping from hazy images to haze-free images.Due to the lack of prior information,the number of parameters and FLOPs(floating point of operations)of such methods are usually large,and the model training faces the risk of overfitting.Moreover,most of them rely on large-scale paired training data which are difficult to collect.The thesis researches how to utilize the haze-prior to solve abovementioned problem.The main contributions can be summarized as follows:Firstly,a haze-prior induced dehazing network(Hi-DehazeNet)is proposed.The dehazing model consists of three parts:haze-prior estimator,dehazing backbone network and post-processor.The haze-prior estimator combines a handcrafted prior with deep neural network to fetch a more robust haze-prior.The haze-prior is used as regularization constraint for the training process of dehazing backbone network,which relieves the problem caused by the absence of pair-wise real training data.The post-processor is employed for further color and detail restoration.Experimental results show that this thesis is comparable with SOTA.In addition,the haze-prior estimator is a plug-and-play module,which can be flexibly instantialized with other existing networks to provide regularization for improving the robustness of dehazing.Secondly,a multiple prior induced dehazing network(MPD)is proposed.The regularization constraints provided by single handcrafted prior are limited.To solve this problem,a multiple haze-prior estimator is proposed,which integrates multiple handcrafted priors into the deep neural network to estimate the haze-prior.Different handcrafted priors make statistics on hazy and haze-free images from different sides,which complements each other and provides more accurate constraint information.Experimental results show that the performance of the multiple prior method is superior to the single prior method.Finally,a multiple prior induced progressive dehazing network(MPPD)is proposed.Existing single-stage image dehazing methods are difficult to simultaneously remove haze of different concentrations.While multi-stage image dehazing methods,which have complicated training process,contain more parameters and consume more computing resource.In order to mitigate this problem,the thesis proposes a feature resuing strategy to reduce the calculation of redundant information,which can reduce the amount of parameters and FLOPs of the progressive dehazing network.The prior-inspired fusion module which utilizes the local max contrast prior is used to calculate the weight of different stage dehazing results from the spatial and stage perspectives.Experimental results show that the parameters and FLOPs of this thesis are much smaller than those of GDN and FFANet.Especially,the MPPD are better than SOTA methods on real hazy dataset.