Ship Detection In High-resolution Optical Remote Sensing Images Based On Deep Learning

Ship detection in optical remote sensing images is of great significance to marine monitoring and territorial security.Currently,with the improvement of the remote sensing technology,more and more high-resolution remote sensing images are that bring about more texture and shape details as well as interference from complex context,such as harbor buildings,waves,and reefs.So far,it has been difficult to overcome the interference of redundant information in existing methods,so that we need an accurate ship detection method which can learn from massive remote sensing images.This paper mainly includes four facets.Firstly,traditional feature descriptors extraction and classification methods are studied Moreover,we study the theory of deep learning and analyze its characters and advantages.Deep learning based methods have deep multilayered architecture and can automatically learn accurate character representation from massive remote sensing images without manual operation.Thus,it is often applied to object detection in complex scenarios.We utilize the typical one-stage detector,single shot multibox detector(SSD),to the ship detection in consideration of its better speed/accuracy trade-off.Secondly,we build a data set and annotate over ten thousand images for ship detection and use deep learning platform Caffe to train SSD model on this data set.The experimental results show that it improves the performance of the traditional method that based on the salient model and LBP(local binary patterns)features for more than 10 percent.Thirdly,we find out that the feature maps used to make predictions contain incomplete information and it is the reason why SSD has the box-in-box problem.Thus,we propose the deconvolutional residual block to generate new feature maps for prediction.This architecture is composed of three branches.Branch 1 and 2 are designed to learn fine-grained and coarse-grained features.Branch 3 deconvolves the feature maps of the consecutive layer to integrate the richer information into the feature maps of the current layer.Then the output feature maps of the three branches are summed as the new feature map.This architecture not only solves the box-in-box problem,but also separate the backbone network from the prediction layers to decrease the interaction between the two parts when updating the gradients.Finally,in the training stage,the number of foreground and background is extremely unbalanced because there are much fewer ship targets than the backgrounds in the images.So we replace the multi-cross entropy loss of original SSD with the binary focal loss to reduce the impact of the detection performance.In our experiments,we compare the detection performance of four detection methods: the tradition method based on the salient model and LBP features,SSD,Faster R-CNN and the modified method with deconvolutional residual blocks and binary focal loss.The precision of these methods on the test set are 81.3%,93.0%,92.8,and 93.8%.,respectively and the results show that our method achieves the highest precision and best speed/accuracy trade-off.