Object Detection Method In Haze Based On Feature Fusion Attention Network

Extracting visual information from the surrounding environment is significant to human activities.However,the normal extraction of visual information can be severely undermined in a variety of complex situations.When the normal extraction of visual information is limited by atmospheric pollution,such as smoke,haze,dust and so on,human activities in nature will be seriously affected.Images collected in the above fuzzy scenes not only seriously influence the execution of computer vision tasks,safety of human but also can even be threatened due to the lack of visibility--such as traffic accidents caused by the inability to clearly detect the road ahead in foggy weather,which increase the necessity of object detection research in the foggy conditionsIn recent years,image dehazing has increasingly become an indispensable topic worth studying in the domain of computer vision tasks.Its purpose is to efficiently restore the basic information of the image from the corrupted input.To provide an effective solution for accurate acquisition of visual information in the real haze environment,this dissertation aims to combine the existing dehazing network with the object detection network on the basis of improving,and verify the feasibility and effectiveness of the proposed way through relevant experiments.The work done is shown as below:Firstly,the image dehazing algorithm is systematically explained,and the characteristics of different dehazing technologies are introduced.Then,a feature attention-based dehazing scheme was developed based on the existing algorithm.The core element of this method is constructing a new self-adaptation feature attention module,which integrates the attention mechanism and the deformable convolutional technique.It is capable of optimizing the structure of the basic network framework of convolutional neural network.Besides,this module is also able to pay more attention to the dense regions of haze,and process different types of complex information adaptively,to some extent could strengthen the capability of the algorithm to dehaze.Then,a multi-step fusion module is proposed and applied into the network,which can adaptively fuse the features between different steps and complement each other to obtain haze-free images.After that,the object detection technique on the basis of Faster R-CNN are used to recognize the output image after dehazing.It has been proved that the recognition rate of images has been significantly improved compared to before dehazing.Finally,to verify the feasibility of the network,the simulation experiment is conducted on the RESIDE dataset,NH-HAZE dataset,Dense-Haze dataset and the real foggy scene,and the results showed that the dehazing effect are satisfactory.