Offshore Ship Detection Based On UAV Aerial Images

With the development of coastal economy,the number and categories of offshore ships have increased dramatically,so the detection and identification of ships can enhance the supervision off the coast.Traditional port monitoring equipment has obvious shortcomings,such as narrow detection range or fixed field of vision,hence the offshore ship detection based on aerial images of UAV has promising application prospects because of good vision field.However,there are still challenges,i.e.,variable scales and orientations of the offshore ships under aerial images,which make it difficult to extract object features for efficient and accurate detection in real-time.The YOLOv5 object detection model has been selected as the basis model to deal with ship detection under aerial images since its good detection speed and accuracy.To cope with the data imbalance of the aerial images,a dynamic data enhancement method is constructed with a loss feedback mechanism,so that the high-resolution input images can be segmented.The segmented images can then be used for training model with reduced calculation load.A YOLOv5-Angle Classification and Modified Pixel intersection over union（YOLOv5-ACMP）is proposed with angle classification module and improved Io U（Intersection over Union）loss to tackle the ship object rotation problem,which can divide the angular scope of the rotation object into different regions for classification and optimizes the pixel-level Io U by angular distance weights,so as to construct a more effective loss function.Furthermore,the Rotation sensitive Score and Angle Channel Switch Align（RSACS-Align）module is proposed based on angle channel switching and feature score map to solve the rotating feature misalignment,which can adjust the feature extraction area and construct a feature score map to build loss weights for more accurate rotating features.As for the variable scale of the ship objects under aerial images,a multi-channel feature pyramid network is constructed with the attention mechanism introduction into the feature extraction network,hence the YOLOv5-Attention based Multi-scale Aerial Rotation Detection Network（YOLOv5-AMARD）is proposed to realize the effective ship object detection of the aerial images.In order to verify the performance of the proposed ship detection method,the object detection models and network module proposed in this paper are tested using the public dataset and the self-built dataset of aerial ship images,while the proposed method is applied to the YOLOv5 model with different configurations for comparative analysis.Based on the YOLOv5 l model,ablation experiments have been carried out on the public dataset for the proposed methods,among which the average detection accuracy of the YOLOv5l-ACMP model,the YOLOv5l-ACMP+RSACS-Align model and the YOLOv5l-AMARD model can be improved by 4.3%,5.6% and 6.2%respectively compared with the YOLOv5 model.In comparison with the model developed in this thesis with the current mainstream models,experiment results demonstrate that the proposed method can achieve good detection accuracy in the ship detection task under aerial images while meet the real-time requirements well.