Research On Active Recognition Technology Of Sonar Image Target

As the only detection device that can work long distances in the ocean,sonar plays a vital role in ocean exploration and provides data support for underwater target recognition.In the marine environment,due to the reflection,refraction and other phenomena of the acoustic signal during the propagation process,it will propagate in different paths.Therefore,the underwater target recognition based on the sonar echo technique has many interferences and results in low accuracy.The target recognition method of sonar images has the characteristics of high resolution,multi-beam and strong real-time performance.Therefore,underwater target detection and recognition based on sonar images is a hot research topic.Most of the traditional sonar image detection algorithms adopt the manual designed feature + classification method,but the effects of such algorithms often depend on the noise characteristics in the specific application background,so the robustness and the generalization ability is poor.With the wide application of sonar technology,in face of huge sonar image datasets,traditional detection algorithms are also difficult to meet the requirements of big data in terms of processing efficiency,performance and intelligence.In this thesis,the object detection methods based on deep learning is improved and extended from natural image to sonar image,in order to achieve high efficiency and precision detection of sonar image.However,deep learning algorithm is data driven,successful detection is impeded by the lack of large annotated sonar images.Manual annotation is not only tedious and time consuming but also demands specialty-oriented knowledge and skills,which are not easily accessible.In order to greatly reduce the cost of data annotation,this thesis develops an active object detection framework for sonar images,it integrates three active learning algorithms for object detection into detection tasks in a continuous fashion,making CNN more amenable to sonar image analysis to dramatically reduce the annotation cost.The results of the experiments illustrate that the proposed active framework with approximately 35% data can achieve competitive results compared to the CNN’s performance using all data.In addition,in order to enhance the anti-noise ability of the detection model in practical applications,the thesis proposes an active object detection framework for sonar images based on generative adversarial network.It generates a super-resolution feature representation for the noise image,uses a discriminator model to guide the generation process,and finally improves the anti-noise ability of the detection model by alternately training the generator and the discriminator model.Compared with several original detection models,the designed network has good detection performance on the data set constructed in this thesis,and shows great robustness,the labeling cost could be reduced by about 50%.