Daytime Sea Fog Recognition By Combining Attention Mechanism With Deeplabv3+

Effective monitoring and forecasting of the distribution and development trend of sea fog will greatly reduce the harm caused by sea fog,which is also an important part of the current smart meteorology research and has a wide application prospect.Based on the application requirements of all-weather sea fog detection and early warning and prediction for satellite remote sensing images,this dissertation is devoted to the research of sea fog segmentation method for daily sea fog images based on deep learning.On the basis of preliminarily completing the collection and processing of sea fog image data set,a deep learning model and detection method for sea fog segmentation combined with attention mechanism and Deeplabv3+ are proposed.The main research content and stage results are as follows.Firstly,the collection and processing of sea fog image data set are carried out in view of the lack of open data set for sea fog segmentation.Based on the research and analysis of existing meteorological satellites,2562 original sea fog images were collected based on the data of Himawari-8 meteorological satellite,and the annotation of the images was completed with the assistance of professionals.To solve the problem of insufficient sea fog image data set,data enhancement processing is carried out based on a variety of different technologies,including classic image rotation and color transformation,as well as advanced Mixup and Swin IR data enhancement technologies,and Mixup technology is improved.A dynamic Mixup data enhancement method based on pixel-level position correlation was proposed.The feasibility and effectiveness of these data enhancement methods are verified by experimental analysis,and the sea fog image data set which can meet the needs of model training is constructed.Secondly,research on sea fog segmentation model based on deep learning is carried out.In this dissertation,the construction,training and validity verification of Deeplabv3+ deep learning model for sea fog image segmentation are firstly carried out.Secondly,the Deeplabv3+ model is optimized in view of the shortcomings of directly using Deeplabv3+ model to carry out sea fog image segmentation.Firstly,the Efficient Channel Attention(ECA)module is introduced into Xception of Deeplabv3+ model to improve the segmentation performance of sea fog image based on Deeplabv3+ model,because the texture of sea fog image is complex and variable.The feature information extracted by the network in the spatial domain and the correlation information obtained in the channel domain are strengthened to improve the feature extraction capability of the backbone network and improve the accuracy of the sea fog image segmentation model.Second,after each cavity convolution of ASPP module of Deeplabv3+ model,CBAM attention mechanism is introduced to adjust the weight share of feature channels through the attention mechanism,so as to solve the problem of weak feature expression of blocked sea fog image,so as to improve the accuracy of image segmentation of the model.The optimized sea fog image segmentation model is called ECA-CBAM-Deeplabv3+ sea fog image segmentation model.The optimized model was constructed and trained,and a comprehensive ablation experiment was carried out on the sea fog image data set,and a full comparison experiment and result analysis were conducted with other image segmentation models.The experimental results showed that the MIo U index number of the optimized ECA-CBAM-Deeplabv3+ sea fog image segmentation model was 85.4%.The index number of MPA is 92.1%,the index of Recall is 84.1%,the index of Precision is 79.7%,and the index of F1 is 81.8%.Compared with the unoptimized model,the index of MIo U is increased by 6.3%,MPA by 5.9%,and Recall by 8.6%.The Precision index increased by 0.8% and the F1 index by 4.6%.This indicates that the optimized ECA-CBAM-Deeplabv3+ sea fog image segmentation model is effective and can provide certain support for business applications.Finally,based on ECA-CBAM-Deeplabv3+ sea fog image segmentation model,the dissertation develops the daytime sea fog image segmentation system.Based on the software development specification,the system needs analysis,database design,overall design and detailed design,as well as the realization of the main functional modules of the system and the test of the system.The system aims to further verify the effectiveness and application effect of the optimized model,and provide real and beneficial information for the improvement and perfection of the subsequent model.