Oil Spill Detection In SAR Images Based On Deep Learning Model

Synthetic Aperture Radar(SAR)has extensive military and civilian value due to its strong penetration ability,omnidirectional,all-weather,large field of view,and high accuracy.It plays an important role in SAR image target detection tasks.In recent years,significant research achievements have been made in the field of image processing using deep learning models and algorithms.This thesis is based on an improved deep learning model to study the method of oil spill detection using SAR image,and the main contents are as follows:1.Given the characteristics of multiplicative noise in SAR images,direct threshold segmentation can result in significant errors.Therefore,this thesis proposes a speckle noise reduction and segmentation algorithm that combines an improved wavelet threshold transformation with threshold segmentation,and applies it to SAR image oil spill detection.First,the logarithmic transformation is used to change the noise property of the SAR image.Then,the processed image is divided into low-frequency and high-frequency sub-images using a three-layer 2D discrete wavelet decomposition,and the sub-images are processed using threshold functions.Finally,the sub-images are reconstructed and subjected to Otsu threshold segmentation.The thesis compares the segmentation results of traditional filtering algorithms with the improved wavelet threshold transformation,demonstrating the effectiveness of the improved wavelet threshold transformation in reducing noise and segmentation.2.Traditional noise reduction and segmentation algorithms that combine image filtering with threshold segmentation are prone to losing target information,unclear segmentation edges,and inaccurate segmentation of oil spills.Therefore,this thesis proposes a feature merged U-Net model(FMU-Net)that combines wavelet threshold transformation,attention mechanism,and U-Net for oil spill detection.Firstly,an improved wavelet threshold transformation and Otsu threshold segmentation are used to perform noise reduction and segmentation on SAR images.Then,the noise-reduced and segmented image and the original image are used as inputs for the U-Net model to perform feature fusion,effectively avoiding data overfitting.Finally,residual network modules are used in the U-Net model to replace traditional convolutional layers to prevent the gradient explosion phenomenon.The SE attention module is integrated into the residual network to improve the model’s ability to extract important features.Moreover,a Coordinated Attention Network(CANet)is introduced at the skip-connection to enhance the network’s ability to extract edge features while suppressing redundant features and reducing the number of model parameters.The proposed model achieves a target detection accuracy of up to 97.02% on SAR oil spill images from different scenarios.Experimental results comparing the proposed model with other models prove its effectiveness.3.To address the issue of insufficient feature extraction,small receptive fields leading to target information loss,and the generation of redundant features in traditional convolutional neural networks,this thesis proposes an ASA-DRNet network model based on axial selfattention modules and Res Net-18 for oil spill detection.This model can extract image features at multiple scales while increasing the receptive field and achieving high performance under certain resource conditions.Firstly,the axial self-attention module(ASA)is used to replace the simple convolution operation in the backbone network(Res Net-18)to enhance the network’s ability to extract low-level features.Then,the ASPP module is optimized through proportional reduction of dilation rates of atrous convolution,replacement of 2D convolution with two concatenated 1D convolutions,and dense connections(feature sharing)among different layers.Finally,feature fusion is achieved by connecting bottom-layer information of different resolutions to strengthen edge features.Experimental results show that the improved network model achieves a maximum recall rate of 76.75% in SAR image target detection,which is significantly better than other networks.This further confirms that the improved network model has high overall performance and is an effective model for SAR image oil spill detection.