The End-to-end Algorithm For Single Image Dehazing Based On Multi-feature Fusion

Image dehazing refers to dehazing the hazy images collected in hazy days to make the images clearer and improve the image quality.In recent years,there have been more and more researches on image dehazing algorithms,and some achievements have been achieved.However,these models still have some problems,such as the balance between model performance and running time,the model cannot effectively remove dense haze,Models trained on synthetic datasets cannot be effectively transferred to real scenes and lack of applicability research in advanced vision tasks.These problems lead to the inability of dehazing algorithms to be well applied in actual scenes.In view of the above problems,the research contents and innovations of this paper are proposed as follows:(1)Aiming at the problem that existing image dehazing algorithms cannot effectively remove haze in dense hazy scenes,and there are haze patches,an end-to-end global feature fusion-based image dehazing network(GFFN)is proposed.In this paper,a multi-scale global context fusion(MGCF)block is designed to capture global context features and integrate them into the network to assist the restoration of dense haze areas;then a simplified pixel attention(SPA)block is used in series with the MGCF block to form global features Fusion Attention(GFFA)module;then based on local residual learning and GFFA module,feature enhancement(FE)module is proposed as a basic module to build an encoder-decoder architecture,enabling the network to focus on enhancing more useful features.The experiments in this paper prove that the algorithm has better performance than the recent state-of-the-art models,outperforming the second-ranked algorithm by 1.13 d B and 0.014 in two evaluation metrics(peak signal-to-noise ratio and structural similarity),respectively,which can effectively eliminate haze patches.It is suitable for dense haze scenes and can achieve fast dehazing.(2)Aiming at the problem that most of the existing image dehazing networks are trained on synthetic datasets and cannot be effectively transferred to real scenes,this paper proposes a An end-to-end image dehazing network(PFFN)for empirical feature fusion.First,this paper designs a priori feature extraction(PFE)module to extract dark channel prior and color attenuation prior features in a way that supports backpropagation.Secondly,this paper designs a prior feature adaptive fusion(PFAF)module that can adaptively fuse two prior features by combining the attention mechanism.Finally,the fused prior features are added to the decoding stage of the backbone network in a multi-scale fusion manner.TThe test results in this paper prove that this model has the highest values for both evaluation indicators on the OTS and O-HAZY datasets,and the peak signal-to-noise ratio is 0.48 d B and 0.44 d B higher than the second-performing algorithm,respectively.The model can better transfer to real scenes.(3)In order to make the dehazing algorithm better serve advanced vision tasks,this paper takes the object detection task as an example,and evaluates and analyzes the application and applicable scenarios of the algorithm proposed in this paper in the object detection task through experiments.GFFN The recovery effect of the global information and the PFFN model is better,and it is more suitable for the scene of object detection of large objects.For small object detection,this paper proposes an end-to-end image dehazing algorithm(ORSBN)based on the original resolution subnet.The feature map in the original resolution subnet keeps the original resolution unchanged,so as to retain more detailed information features,making the dehazed image more suitable for small object detection.In this paper,three image dehazing models are proposed for different problems.The parameters of the three algorithms are all within 10 M to ensure the efficiency of the model,and the effectiveness and advancement of the three models are proved in the object detection task.