Ship Detection Via Convolutional Neural Network

In recent years,there have been an increasing number of conflicts concerning marine interests in the world,ship as the most important carrier and military target in maritime operations.Ship detection and recognition not only has a wide range of practical applications in the civilian field,but also has important strategic significance in the military field.This paper studies the ship detection,we use convolutional neural network to detect the ship's target,the details are as follows:First,we propose a deformable Faster R-CNN detection framework for ship detection.We have modified the resnet-101 network structure,mainly includes replacing the conventional convolution with deformable convolution/dilated convolution and adjusting the convolutional kernel stride of the last convolution module on ResNet-101 network.It improves the resolution of the output feature map.We also have adjusted the position of the RoI-Wise subnetwork in the detection framework,modified the full connection layer in the RoI-Wise subnetwork,and replaced RoI Pooling with deformable RoI Pooling.At the same time,we also modify the loss function of deformable Faster R-CNN and replace it with Focal Loss.Finally,we use the Decay-NMS for post-processing.The experimental results show that the changes in our thesis have a good effect on small target detection.Secondly,a deformable R-FCN algorithm is proposed for ship detection.We have modified RoI Pooling Layer,loss function,and post-processing method,it including the normal convolution to deformable convolution/dilated convolution,deformable PS RoI Pooling,Focal Loss and Decay-NMS.It improves the mAP of the ship detection.Thirdly,we propose a deformable FPN detection framework for ship targets.Convolution kernels are constructed for the convolution layer of the feature pyramid network.The original RoI Pooling Layer is replaced by deformable PS Rol Pooling,and the loss function is modified with Focal Loss.In the end,we use Decay-NMS for post-processing.Comparison experiments verify the effectiveness of the deformable FPN.