| Underwater target recognition is one core technology of underwater unmanned exploration,which is very important for marine resource development and marine defense.Since sonar image can visually present a target in muddy water and at a long distance,it becomes an important medium for underwater automatic target recognition.However,the complex marine environment,the variety of targets and the low image quality bring great challenges to the sonar image recognition problem.Traditional underwater target recognition methods rely on experience to extract image features and use shallow classifiers to classify targets,which leads to low recognition accuracy and limited generalization ability.The research on the recognition method with stronger feature characterization ability and generalization ability can improve the underwater target recognition accuracy,which is the only way to realize the automation and intelligence for underwater unmanned platforms.In this paper,deep learning theory and methods are closely combined with underwater target recognition applications,to improve the accuracy of sonar image recognition.Research on underwater target sonar image recognition based on deep learning is carried out,main research contents and innovations include:1.A real measured sonar image data set which can be used for classifier training and testing is established.In view of the lack of data in the field of sonar image recognition,through extensive exploration and collection,a total of 2915 actual sonar images,including ship wreckage,aircraft wreckage,underwater robot,tank wreckage,engineering foundation stone,and diver,are obtained.The acquired sonar images are manually classified according to the target category in the image,and each image is manually labeled with a category label to establish a sonar image data set contains 9 classes of underwater targets.The number of images for each class ranges from 195 to 583.Underwater targets in the sonar images have different sizes and different angles of view,which can fully test the performance of image noise reduction and underwater target recognition algorithms in practical applications.2.A fast noise reduction method for sonar image is proposed.The human visual attention mechanism usually guides the eyes to salient region and gives priority to that visual information.In view of this,saliency detection based on the manifold ranking is introduced into sonar image processing,to divide the image into two parts: salient region and non-significant region.For the salient region with small proportion,the BM3 D algorithm is adopted to reduce noise and protect the main information of the image;for the non-significant background which is not very concerned,high efficiency mean filtering is used.On the established sonar image data set,the proposed algorithm is compared with the classic MF and BM3 D algorithms through the subjective and 2 objective evaluation indexes.The experimental results show that the efficiency of the proposed algorithm is much higher than that of BM3 D,while the image visual effect is guaranteed,which can better meet the application requirements of autonomous underwater vehicle.3.An underwater target recognition model based on convolutional neural network with spatial pyramid pooling and significant region segmentation is proposed.Firstly,the sonar image is segmented and clipped with a saliency detection method to reduce the dimension of input data,and to reduce the interference of image background to the feature extraction process.Secondly,by using stacked convolutional layers and pooling layers,the high-level semantic information of the target is automatically learned from the input sonar image,to avoid damaging the effective information caused by extracting image features manually.Finally,the spatial pyramid pooling method is used to extract the multi-scale information from the sonar feature maps,which is to make up for the lack of detailed information of sonar images and solve the problem caused by the inconsistent size of input images.On the collected sonar image dataset,experimental results show that the target recognition accuracy of the proposed method is better than that of the traditional recognition methods.Besides,the designed model can recognize underwater targets more accurately and efficiently than conventional convolutional neural networks.4.Underwater target recognition in small sample size situation based on transfer learning with deep convolutional neural network is proposed.In order to solve the contradiction between insufficient sonar image samples and the need to further strengthen the network feature extraction ability,the transfer learning method is studied and applied to sonar image recognition.An end-to-end DCNNs model,named Echo Net,is designed for sonar image feature extraction,and a network training method based on transfer learning is developed.Experimental results demonstrated that our method can implement efficiently,and the recognition accuracy on a nine-class underwater automatic target recognition task reached97.3%,outperforming traditional feature-based methods.Besides,the effect of the number of frozen layers on the recognition performance is analyzed,and the reason why the optical image knowledge can be successfully transferred to the sonar image recognition is discussed.5.An underwater target class incremental learning method based on double weight consolidation is proposed.In order to overcome the catastrophic forgetting problem of recognition model in the process of continuous learning,and to make the underwater target recognition system adapt to the dynamic characteristics of the target category in the real world,a double consolidation class-incremental learning method is proposed.In the process of incremental learning,the network model parameters are adjusted by means of knowledge distillation and elastic weight consolidation,so that the network retains the ability to recognize the old category targets in the process of learning new category knowledge.The incremental learning experiment is designed,and the proposed method is compared with the popular incremental learning methods such as EWC,Lw F and i Ca RL.Experiment results show that the proposed DCCIL method takes advantage of the knowledge distillation,the weight consolidation and nearest neighbor classifier,the recognition accuracy is better than that of the current popular incremental learning algorithms,and close to the performance upper bound.The proposed method can effectively improve the expansibility and intelligence of the recognition model.In this paper,sonar image recognition technology based on deep learning is deeply studied and discussed.The deep learning model is designed to automatically extract implicit features of the sonar images,and at the same time,the learning ability of the recognition model is enhanced in the case of small sample and target classes increased.The proposed methods greatly improve the accuracy and intelligence level of underwater target recognition technology. |
