Research On Feature Extraction Of Side Scan Sonar Images Based On Deep Learning

There are some questions in sonar images features extract,such as strong speckle noise,low image resolution,poor image quality,and difficult target segmentation.In order to overcome the shortcomings of traditional algorithms in sonar image feature extraction,we apply instance segmentation network to sonar image feature extraction.So we use Mask RCNN instance segmentation network to extract image features.The effectiveness of the deep learning algorithm is verified by experiments.Firstly,we build required sonar image feature extraction data set for Mask RCNN instance segmentation network training.Then,according to the working principle of side-scan sonar,we parse sonar echo map and obtain real submarine topographic image.According to the noise analysis of the side scan sonar image,we add Rayleigh distribution noise and affine transformation to the sonar image set to expand dataset size.More than that,we use Labelme software to make image label data of Mask RCNN network training and testing.According to the set proportion,dataset is divided into training set,verification set and test set;Secondly,under the condition of pre-training model models,we use Mask RCNN instance segmentation network respectively based on ResNetvl and ResNetv2 feature extraction network to experiments.The results of the two experiments were compared,and the better model was selected for subsequent training.Due to the limitations of experimental equipment,Batch Normalization can only be set to small value.However,Batch Normalization can cause a large error under small batch size.Therefore,we replace Batch Normalization with Group Normalization which is insensitive to batch size.At the same time,in order to ensure that the negative input is not completely truncated like Relu activation function,but allows relatively small negative gradient inflow to ensure the smoothness of every point,Mish activation function is introduced to replace Relu activation function to improve network performance;Then,based on the previously trained model,we imitate ResNeXt and Res2Net convolutional neural network adding grouping convolution to the network.Then according to IResNet convolutional neural network to improve the skip layer structure to avoid unnecessary information loss.At the same time,we add dual-channel attention mechanism in the different ways and obtained better model which have a better feature extraction capability.Furthermore,we do ablation experiments aiming to solve the problem of residual network redundancy and compare the experimental results.On this basis,we replace Feature Pyramid Network(FPN)with bidirectional cross-scale connections and weighted feature fusion(BiFPN)which add additional weight to each input feature map during feature fusion to let the network learning the importance of different sizes features;Finally,due to the great difference in the proportion of foreground and background in the side-scan sonar data set,we innovatively use Focal Loss function to replace mask loss function to improve the performance of network semantic segmentation.At the same time,we replace Intersection over Union(IoU)with Complete Intersection over Union(CIoU)which make regression loss function take overlapping area into account,center distance and aspect ratio.Then,the NMS algorithm is improved to DIoU-NMS algorithm,which not only considers the overlapping region,but also considers the center point distance between the two boxes.The warm restart cosine annealing algorithm is used to improve the learning rate of the network and avoid network training falling into the local minimum.After the model is trained,the interface of image feature extraction of side scan sonar is designed.