Marine Ship Detection And Direction Prediction Based On Convolutional Neural Network

The further development of remote sensing technology has made high-resolution remote sensing images easily available.For a long time,ship automatic detection has played an important role in the field of remote sensing,and ship parking information and direction information have important research significance.The development of ship inspection technology is not only conducive to strengthening maritime supervision,but also reflects the scientific and technological strength of a country.Ship remote sensing detection is an important subject in the field of remote sensing,and it is also a challenging subject.The complexity of application scenarios,the redundancy of detection areas and the intensive detection of ships are the main obstacles that restrict the successful operation of traditional methods in remote sensing ship detection.At present,ship detection mostly relies on traditional detection methods and machine learning methods.Traditional methods cannot make full use of the rich spectral information and spatial information of remote sensing images and may ignore the location information of the target.In addition,because the target size in remote sensing images is too small And the distribution is dense,so the use of traditional algorithms is easy to cause missed detection of the target,which reduces the detection accuracy.In response to the above problems,this paper improves the traditional FASTER RCNN network and proposes a ship detection and direction prediction method based on convolutional neural network.The main work is as follows:First of all,in view of the problem that the size of the target in the remote sensing ship image is too small to cause missed detection,this paper uses a multi-scale feature pyramid network,which can fuse multi-layer features to enhance the expressive ability of features,thereby improving the detection accuracy of small targets rate.Secondly,in view of the wide distribution of ship targets in remote sensing images,this paper sets anchor frames of different scales in each layer of the feature pyramid.According to the characteristics of the ship’s length-to-width ratio,the anchor frames of the same scale are set to 9 types of lengths.The aspect ratio enhances the adaptability of the network to ships of different sizes.Then,in order to address the problem of the cluttered distribution of targets in remote sensing ship images,this paper proposes a detection method based on rotated rectangular frames,which defines the direction information based on the traditional horizontal detection frames,and proposes a rotated Ro I alignment method and rotated non-maximum suppression,and the final network output is a bounding box with direction information.The rotated bounding box not only improves the detection effect of the model,but also preserves the position information of the target,so it is more suitable for the detection of densely distributed and directional targets.Finally,this article uses the idea of migration learning to solve the problem of lack of pictures in the data set of this article.Based on the DOTA data set and self-built data set,it finally realizes the ship detection and direction prediction based on the convolutional neural network,and combines it with the traditional deep learning detection The method compares the detection effect,so that the method in this article is the most advantageous.