Image Segmentation Of Irregular Drift Ice In SAR Images Based On BiSeNet

Sea ice has a great impact on climate change and the safety of navigation at sea,so the real-time prediction of sea ice can effectively reduce the impact of sea ice on production practices.The characteristics of sea water surface covered with melting ice,drift ice and melting ice texture are very similar,and the characteristic discrimination between them is low,resulting in small differences between sea water and drift ice,and the sea water covered with melting ice is easy to be mistakenly divided into drift ice;Due to the influence of illumination,radar echo intensity and other factors,the brightness,darkness and roughness of ice are different,resulting in large intra-class differences between thin and thick ice;at the same time,the SAR sea ice image is seriously disturbed by speckle noise,with limited spatial resolution and lack of features,which reduces the accuracy of ice segmentation in SAR image.The accuracy of sea ice image segmentation will directly affect the extraction of sea ice position and area information,and sea ice detection also needs to be fast and real-time,so the accuracy is the key to sea ice segmentation,and the speed is also a factor that can not be ignored.In order to solve the above problems,this thesis proposes a method for irregular ice segmentation in SAR images based on dual-branch network structure.The main work and research contents of this thesis are as follows:Firstly,the SAR sea ice data is preprocessed and the data set is made,in this thesis,the data is clipped from the three SIR-C polarimetric SAR sea ice image data source,and the effective samples are selected for data enhancement.In order to enhance the experimental effect,the data noise processing is carried out,and in order to reduce the over-fitting phenomenon,the brightness of the samples is enhanced.Contrast enhancement and horizontal flip,random rotation operation to increase the amount of data,finally get a data set containing3322 samples,effectively improve the detection effect of the model.Secondly,in order to obtain rich spatial information and receptive field at the same time,a sea ice segmentation method based on improved BiSeNet is proposed.There are three convolution layers in the spatial branch,which mainly capture the low-level spatial details to get high-resolution feature representation,and mainly obtain the details of sea ice color,edge and so on;In the context branch,the pre-trained classical network Res Net is used to extract deep features for semantic classification.After the branch,the global average is used to stabilize the large receptive field,which reduces the amount of calculation.The parallel calculation of the two branches improves the operation speed of the model.Thirdly,based on the original BiSeNet network,this thesis adds a jump structure,which can effectively combat the high noise of SAR image and the lack of shallow feature information,and the segmentation performance has been significantly improved.In the process of network training,through the test and comparison of different learning rates,the optimal learning rate is obtained for model training.Finally,the improved bisenet network is applied to thousands of preprocessed sea ice images to train the segmentation model,and the method is compared with the classical segmentation algorithm(FCN)and the lightweight network model(ICNet)proposed in recent years.The experimental results show that the comprehensive performance of this method is better than the other two models.