Research On Single Image Dehazing Algorithm Based On Deep Learning

The presence of haze not only causes limited visibility in outside situations,but it also degrades the contrast and sharpness of the outdoor obtained images,which has a negative impact on a range of subsequent advanced vision tasks.Hence,the development of efficient single image dehazing algorithms is a difficult and extensive task.Most physical modelbased dehazing algorithms forecast transmittance and atmospheric light values,and any mistakes result in unsatisfactory dehazing results.While the end-to-end dehazing methods can immediately obtain a clear image,there are frequent issues such as detail loss and color distortion.For the aforementioned issues,this paper employs deep learning-related techniques to conduct the single image dehazing research described below:(1)We propose a topological network for multi-scale feature extraction based on deep learning,concentrating on the acquisition of rich images features at diverse scales with hierarchical extraction and fusion.In the feature extraction stage of the network,a residual learning is introduced,and features are transferred across diverse scales on the jump connections.In the feature fusion stage,an attention mechanism is introduced with two modules:Adaptive Feature Selection Unit and Feature Fusion Unit.We adopt them to adaptively elect and fuse significant image features at diverse scales,discard redundant image features,and reach better image dehazing effectiveness.In addition,the performance of the network architecture is significantly improved and higher accuracy is attained.Compared with the current sophisticated dehazing algorithms,the proposed method has remarkable benefits in both quantitative and qualitative analysis.(2)We propose a densely integrated multi-branch attentive network with a multibranch structure as its backbone that is adequate to extract contextual information.The network consists of two parts: a multi-scale feature extraction module and a dense-feature fusion module.They contain efficient dehazing modules: residual channel attention blocks,a multi-level feature fusion module for front and back layers,a dynamic select feature module,a color information renovation module,and a feature enhancement module.The feature maps can be self-corrected to refine the feature information at diverse levels via a succession of feature extraction,interaction,and fusion.The ablation experiments demonstrate the effectiveness of the modules that form the network and can visibly enhance the image quality.Compared with other dehazing algorithms,this method provides better objective、subjective evaluation results and can alleviate color distortion and incomplete dehazing.It also strikes an excellent balance among model size,execution time and performance.