Research On Ship Recognition Algorithm Based On Deep Learning

Ship recognition technology has broad application prospects in water transportation management,marine resource exploration,disaster rescue and anti-smuggling.Compared with synthetic aperture radar images,infrared images and remote sensing images,there are more detail features such as color and texture in visible light images,which has advantages on real-time ship detection and fine-grained classification.At present,traditional ship detection algorithms have low detection accuracy and poor robustness under the conditions of variable light,cloud occlusion,complex coastal backgrounds and multi-scale ship targets.There are relatively few researches on fine-grained classification of ship targets.In view of the above deficiencies,this paper conducts the following researches on ship recognition based on deep learning methods:(1)In order to improve the generalization ability of the model on the small-scale ship dataset,this paper makes data augmentation for ship images based on the improved Deep Convolutional Generative Adversarial Networks(DCGAN).In view of the low resolution and blurry details of ship images generated by DCGAN,this paper proposes three improved strategies of optimizing the network structure of the generator and discriminator,using the Non-local module to capture global information and adding feature matching loss to generate more clear and realistic ship images.Experiments show that the ship data augmentation based on the improved DCGAN algorithm in this paper can effectively improve the accuracy of ship identification.(2)In order to improve the accuracy of ship detection,the ship image is prepocessed based on the improved dark channel prior algorithm first.Then this paper analyzes the principle and model structure of YOLOv2 in depth and makes several improvements.Multi-scale feature fusion method is used to replace feature reorganization fusion method of YOLOv2.Anchor box that conforms to the shape of ship is designed based on K-menas clustering.The loss function is improved based on Focal loss for hard example mining and relative width-to-height deviation.Experiments prove that the algorithm in this paper can effectively improve the detection accuracy of multi-scale ship targets under complex background.(3)In order to improve the accuracy of ship fine-grained classification,the PAM-CAM attention module is proposed by combining spatial and channel attention mechanism,which is embedded in the improved VGG16 network.The cascade classification network is further designed:the global image and the salient local region of the ship extracted by the Grad-CAM algorithm are used as the input of the object-level classification sub-network and the local-level classification sub-network respectively,and the global features and local features are fused to realize the fine-grained classification of ship targets.Experiments proves that the attention mechanism helps the model to focus on the local key regions,and the addition of local detail features can effectively improve the classification accuracy.Experimental results show that the algorithm in this paper improves ship target detection accuracy and ship fine-grained classification accuracy,which effectively realizes ship target recognition in complex scenarios.