Deep Learning Based SAR Image Ship Target Detection

In the current global situation,the deployment of marine strategy has become a top priority,China’s territory and territorial waters are vast and far,and rich in various marine resources,with the construction of China’s maritime power,maritime ship target detection is not only widely used in the civilian field,but also plays a crucial role in the military field,Therefore,the use of synthetic aperture radar(SAR)images for maritime ship target detection has received great attention from all parties.This thesis firstly applies rotating anchor frame target detection to SAR image detection,and for the first time uses S2A-Net(Single-shot Alignment Network),Faster R-CNN,Ro I Transformer,FCOS and Retina Net algorithms from the latest JDet remote sensing image detection algorithm library combined with satellite remote sensing big data,Using the SAR image dataset SSDD+ with rotating border annotation,we conducted a study on the target detection of maritime ships and achieved the target detection support for SSDD+ dataset.We learned and understood various algorithms in JDet algorithm library and compared them experimentally,among which the S2A-Net result for rotating anchor frame detection was the most superior,and its map(mean average precision)reached 89.8%,and among the detection results of SSDD+ by major network frameworks,this framework can achieve higher precision in remote sensing image detection of rotating anchor frames.Secondly,the latest Swin-transformer model is combined with satellite remote sensing big data,and in order to pursue higher detection accuracy,an improved version of Augmix method,Multiple-augmix,is proposed to achieve target detection and data enhancement of the dataset at the same time,and the experiments show that the Swin-transformer algorithm The Swin-transformer algorithm is suitable for remote sensing maritime ship target detection,and the detection accuracy of Swin-transformer can reach 86.5% for the original SSDD dataset,and 96.2% for the enhanced SSDD dataset using Multiple-augmix method,which is 10 percentage points higher than that before enhancement.From the perspective of network model,this paper cleverly uses the self-attention mechanism to take advantage of its image detection features and adds it to the network framework of YOLO-v3.Comparing with the original network,the latest network performs excellently,achieving not only improvement in detection accuracy but also further optimization in detection speed.Finally,in the field of target detection,in addition to the processing of the algorithm framework,there are many mature data enhancement methods to optimize the quality of the "data",but not every Augmentation method can be applied to SAR image maritime ship target detection,in order to address this problem,this paper selected five major SAR image datasets.In order to address this problem,eight data Augmentation methods are selected,eight data Augmentation methods are applied to each of the five major datasets,and then four detection models are selected to conduct "one-to-one pairing" detection experiments with 45 datasets(original + enhanced),generating a total of 180 sets of detection results,proving through statistical experiments that the data Augmentation of maritime SAR image datasets is The optimal use of each data Augmentation method in the SAR image dataset is also summarized.