Automatic Target Recognition Of Sonar Image Based On Weakly Supervised Localization

Since the 21 st century,countries all over the world have begun to pay more attention to the development and utilization of marine resources.China has also put forward the new marine strategy.As the key technology of marine development and utilization,marine environment perception system has been paid more and more attention.Compared with the strict requirements of optical sensing devices for lighting conditions,sonar sensing technology is more used in the construction of marine environment sensing systems because it is more suitable for an underwater environment.However,the localization and recognition technology for sonar image target has been blocked by the lack of open source sonar image datasets.Most of the classical methods solve the task of location and recognition separately for a specific dataset.These methods require a lot of labor cost and lack of generalization ability.In this paper,a sonar image automatic target recognition technology based on weakly-supervised localization is proposed.By introducing the Grad-CAM target visualization technology into the classification convolution neural network,the sonar image target localization and recognition only rely on the sample category label is realized,which greatly reduces the labor cost.At the same time,the over-fitting problem of training models caused by the lack of sonar datasets is also solved.This paper optimizes the sonar image automatic target recognition technology based on weaklysupervised localization from aspects of data parameters,network loss,and backbone model to improve the accuracy and robustness.For the over-fitting problem,it is a common method to pre-train the model on a larger dataset to learn extracting features.In view of the great difference between optical dataset and sonar dataset,we first try to find a suitable dataset to obtain the pre-training parameters which are helpful for sonar images.Specifically,the Ada IN style-transfer method is used to transfer the optical dataset into the dataset based on shape preference,so that the model will focus on improving the ability of shape feature extraction in the pre-training process.With the help of special pre-training parameters,sonar image target location,and recognition accuracy is greatly improved.Because the pre-training parameters are easy to be destroyed by low-quality samples in training,we further seek to optimize our technology from the network loss aspect and propose a deep shape preference network(DSPN)based on domain adaptation.The network uses MK-MMD to measure the distribution distance between the sonar image dataset's features and another auxiliary dataset's features which mainly contain shape features,and then we introduce the distribution difference as an additional constraint into the loss optimization in the training progress,so that the model can select effective shape features to guide the classification and localization in the training process.Thus,the over-fitting problem in the process of model training is greatly alleviated.In order to further verify the robustness of the DSSN module,and to find a better method to solve the task of automatic target recognition in sonar image on the backbone model aspect,we replace the Res Net-18 backbone model with several more advanced backbone models.Through comparative experiments,we verify the reliability of the proposed ”pre-training parameters + DSPN module” method.After using Dense Net-121 as the backbone model,the optimal automatic target recognition result of the sonar image is obtained.At the same time,it also explores the possible direction to further improve our technology.