Research On Object Detection In Visible Images For Ship Intelligent Navigation

With the rapid development of sensor technology,network communication,information processing,artificial intelligence and so on,as well as economic and security considerations,the global shipping industry is actively exploring the way to decrease the number of crew members and improve the level of ship automation and ship intelligent.The intelligent navigation technology of ships,which meets the future development demand of shipping industry,has become a new field which the world’s major ocean countries are striving to develop vigorously.Intellisense uses the sensor system on the intelligent ship to detect the objects in ship navigation,which is the information base and one of the key technologies for intelligent navigation of ships.As one of the main sensors to realize intellisense,visual sensor has the advantages of rich information,simple layout and low cost,so it is one of the main perceptive approaches to realize ship intelligent navigation.This paper mainly focus on the visible image object detection method for ship intelligent navigation and conducts a in-depth research.The main research work includes:(1)A few-shot learning method is designed to solve the problem of insufficient training samples for object detection model caused by the difficulty and huge time-consuming of water surface visible image collection and manual annotation.First,a large universal image dataset which contains water surface images is used as the source domain to train the network model,and the model weights are sent to the model trained in the target domain for network-based deep transfer learning;Then,combined with the environmental characteristics of ship sailing at sea,appearance consistent heatmap guided poisson copy method is used to well-directed copy the objects on the image of the target domain,which increases the diversity of image samples.An experiment on images taken from real sea areas shows that the training effect of object detection model under the condition of insufficient samples has been obviously improved by the proposed method.(2)Aiming at the problems that it’s difficult to extract effective object features as the ship navigation scenes are diverse and the water surface visible images acquired by onboard visual sensors are complex,a strong semantic feature extraction method for surface image object detection is proposed.First,deformation convolution is uesd in the backbone network to make the convolution sampling points have translations to adapt to the geometric transformation of object and obtain adaptive receptive field.Then,the feature recombination based on semantic information is used in the feature pyramid to adaptively aggregate specific object information through global semantic information,so that strong semantic feature map can be output to improve the accuracy of subsequent object detection.An object detection experiment of real sea area image is conducted on the proposed method and the baseline method,which shows that the proposed method not only improves the accuracy of the object detection method,but also makes the output feature maps of the proposed method clearer after visualization.(3)Aiming at the problem that the scale of obstacles in visible images are variable due to the complex ship navigation scenes,the adaptive feature fusion module is designed,which through a set of learnable weight coefficient to adaptively fuse the feature maps from different size on channels and make the object detection algorithm has a good detection effect on objects from different scale.Meanwhile,the adaptive feature fusion module is applied to Cascade RCNN(Cascade Region-based Convolution Neural Networks)algorithm with high accuracy,and a multi-scale object detection algorithm is designed by introducing soft non-maximum suppression,complete intersection of union loss function and learning rate cosine annealing method.The real sea image dataset is used to represent the open water scene,the Yangtze river waterway image dataset is uesd to represents the narrow water scene,An object detection experiment on both datasets shows that the multi-scale object detection algorithm has good detection effect on objects from large,medium and small scale.