Research On Deep Learning-based Ship Detection

At present,deep learning has made breakthroughs in many fields.The current research and application of deep learning in ship detetion is still relatively little.One of the difficulties in ship detection is that there is no publicly available dataset of the ships.Another problem is that current methods for the ship detection are mostly traditional detection methods,such as HOG + SVM,DPM,etc.The drawbacks of the traditional detection methods are not robust to light,morphological changes,occlusion,etc,and their accuracy on the ship detection is far less accurate than deep learning methods which are gaining more and more attention in recent years.To solve these two problems above,the main work and innovation of this thesis are as follows:1.Related to the difficulty of the samples,a dataset that contains 8526 ship images(marked)is set up in this experiment.Among these ship images,there are 2458 ship images can be divided into nine categories including warships and civil ships,etc.In addition,in order to study the detection performance of small targets in complex background and compare the experimental results in this article with the published experimental results,this thesis runs the test in the ship test set in the VOC2007 datasets and extracts 6020 ship samples from the data in ImageNet2014,VOC2007 and VOC2012 datasets which marked as 'ship' as expansion data.The data in this article contains images of the occlusion,light changes on ships,and ships as small targets.2.In this thesis,three advanced deep learning frameworks for object detection are first proposed to solve ship detection problems.The frameworks are Faster R-CNN,R-FCN and SSD.In the manually established datasets,these three methods have good detecting results when being applied on different categories of ships that have been given in this thesis.The detection accuracy is about 90%,which is 30%higher than that of HOG+SVM.In the VOC2007 test set,they are better than the original test results.Among these results,the results of Faster R-CNN detection is 50.17%higher than that of DPM;R-FCN method which uses depth residual network(ResNet-101),has improved the accuracy of VOC2007 test set to 75.15%,which is 5.52%higher than the best results of Faster R-CNN,and it can detect more small targets.The accuracy of the detection in SSD method is 77.18%and its detection boxes are more accurate,and the detecting speed is up to 40 frames per second,but the number of small targets detected is less than that of R-FCN.The experimental results show that Faster R-CNN is much high accurate than traditional method,but the detecting accuracy and real-time are worse than that of SSD and R-FCN.R-FCN has the largest number of small targets detected in complex environment,which detecting ability in small target is the best among these three methods.SSD method is the best in detecting real-time performance.