Fully Convolution Neural Network Based Ocean Scene Segmentation

In recent years,with the development of graphics processing unit and other hardware devices,deep learning has also been developed rapidly.It has been widely used in various research fields of computer vision,such as image classification,image detection,image semantic segmentation,etc.A large number of practices and studies show that deep convolution neural network can accurately and efficiently complete computer vision tasks.In the field of remote sensing,marine ship image segmentation is an important research direction.In order to obtain the position,course,category and other information of the target ship,it is necessary to segment it quickly and accurately,which provides the basis for further information extraction.This paper mainly focuses on the optical remote sensing image.First,U-Net is improved to complete the task of sea land segmentation.Then,based on the full convolution neural network,a semantic segmentation network model is constructed for the ocean ship.On this basis,the improved optimization realizes the example segmentation of the ship.The main work is as follows:(1)The U-Net is improved,and the skip connection between encoder and decoder is redesigned.The more complex dense connection can better transfer features,fully integrate deep features and shallow features,improve the segmentation accuracy of the network,and complete the sea land segmentation.(2)Based on the fully convolution neural network,a fully convolution network structure of encoder decoder is proposed.The convolution block with residuals is used as the encoder to extract image features,the continuous transposed convolution is used as the decoder,and each level features are effectively integrated between the encoder and the decoder.At the same time,a weighted loss function is used to realize the maritime remote sensing image whole scene semantic segmentation from end to end.(3)On the basis of the above work,according to the multi-scale characteristics of remote sensing image,the feature pyramid network is introduced,and the whole network is modified into a multi task form with two branches,that is,after a shared convolution layer,one branch completes the ship detection task,the other branch carries out the ship semantic segmentation.And finally integrates the information of the two branches to realize the ship instance segmentation.