Ship Detection And Identification Based On Deep Learning

The ocean has received more and more attention in the 21 st century.Whether it is the development of marine resources,the safety of protection operations,the awakening of marine territorial awareness,and the protection of marine territories from infringement,it is necessary to quickly obtain information on maritime targets.How to use modern technology to achieve rapid identification and location of maritime targets has become a very urgent need.With the development of modern imaging technology,the resolution and quality of images are getting better and better.At the same time,with the rapid development of deep learning,it shows its great advantages to traditional methods in the field of vision.Therefore,the method based on deep learning is used to solve the sea.Ship detection problems have become a promising approach.In the field of target detection,it is often encountered that the size of the detection target is different,and the detection target is too small,which brings difficulty in detection.At the same time,in the face of specific detection scenarios,the amount of data trained in the neural network is often insufficient,which will lead to poor final detection results,and it is urgent to solve such small sample training problems.At the same time,when the model is deployed to the mobile terminal,the scale of the model and the actual computing resources of the mobile hardware are weighed.Therefore,in view of the above problems,this paper studies from the following aspects:Firstly,it studies the principle of deep learning,especially the excellent convolutional neural network in the field of image processing,analyzes the mechanism of its automatic extraction of features,studies the target detection algorithm based on deep learning,and analyzes the characteristics and advantages and disadvantages of each algorithm.The corresponding algorithm is introduced into the field of ship detection.According to the problems faced by ship detection: it has good real-time performance,good performance in small target detection,and good intensive detection of target distribution.It is based on convolutional neural network.YOLO3 algorithm.Then,for the specific detection tasks,data collection difficulties often occur,and the amount of data is insufficient.The training of the depth model is often under-fitting,and the detection accuracy is not high.For this small sample training problem,this paper introduces the generation-oriented confrontation network into the ship.In the ship detection,data enhancement is performed on the collected data set to augment the data set to alleviate the under-fitting and improve the effect of the algorithm.Finally,ship detection and identification often requires good real-time performance,and is often used on mobile devices such as small drones,which are often designed to be mobile and lightweight,and do not provide excessive computing resources.Like the hardware support with good computing performance on the server,this article deploys the trained model in JETSON NANO,a new embedded mobile development board to test its performance on the mobile side.In this paper,the detection of the detection target is too small and the distribution is too dense.At the same time,in the scenario where the data volume is insufficient,the data set is expanded by using the generation-oriented confrontation network,and the target of improving the detection precision is realized.When deploying to the mobile terminal,it faces the problem of insufficient computing resources.It uses a streamlined model to deploy to complete the inspection task,evaluates the running time and accuracy,and lays a foundation for the subsequent work.