A Sea Image Detection Model Based On Attention Mechanism And Deep Learning

Intelligent meteorology is a clear goal in the "Outline for High-Quality Development of Meteorology(2022-2035)".Intelligent monitoring of all-weather sea fog and accurate early warning and forecasting services play an important supporting role in the construction of smart meteorology.In response to the need for image-based sea fog detection based on artificial intelligence methods for smart meteorology,this thesis aims to study a daytime sea fog image detection model based on deep learning,proposing a combined attention mechanism and SwinTransformer model and detection method for daytime sea fog detection.Through comparative experiments with different models conducted on the collected and organized real sea fog image dataset,the proposed daytime sea fog detection model achieved the expected detection effect.The main research content and interim achievements of this thesis are as follows.Firstly,in response to the current lack of daytime foggy image datasets for deep learning,this thesis conducted data collection and processing work on foggy image data.To address the issue of insufficient foggy image datasets,data augmentation was performed using techniques such as Augmentor,Mixup,Mosaic,and random jitter,as well as research on data augmentation based on generative adversarial networks.The feasibility and effectiveness of these data augmentation methods were experimentally analyzed and verified,and a foggy image dataset that meets the needs of model training was constructed.Secondly,based on the aforementioned dataset of sea fog images,a deep learning model for detecting sea fog images was studied.The thesis first conducted research on the construction,training,and effectiveness validation of the Swin-Transformer deep learning model for sea fog image detection.Secondly,in response to the issue of smaller receptive fields in the Swin-Transformer-based sea fog detection deep learning model,the thesis proposed the use of a mixed atrous convolution to improve the Swin-Transformer model.Further improvements were made to the feature extraction of the model by first using a feature fusion structure based on weight coefficients,then replacing the self-attention mechanism in the Swin-Transformer with a mixed attention mechanism to optimize the feature extraction method.Finally,to fully utilize the time series information of the sea fog image data,the CLSTM(Convolutional Long Short-Term Memory)was incorporated with Swin-Transformer to propose a sea fog image classification model based on temporal information(denoted as ST-CLSTM)to improve the accuracy of sea fog image detection.The improved model was constructed and trained,and compared with other detection models on the sea fog image dataset through experiments and result analysis.The experimental results verified that the improved model achieved significant improvement in the AUC index over the unimproved model,increasing from 0.823 to 0.996.Finally,based on an improved deep learning model for sea fog image detection,this thesis carries out the research and development of a daytime sea fog image detection system,which recognizes whether there is sea fog in the image.The system follows the principles of software engineering and includes requirements analysis,overall design,detailed design and implementation,as well as functional testing.The test results meet expectations.The developed system aims to verify the practical application effect of the model and provide valuable information for further improving the sea fog image detection model.