Research On Ship Detection Algorithm In Optical Remote Sensing Image Based On Deep Learning

Modern remote sensing technology can quickly and accurately provide Earth observation data and obtain high-resolution optical images.Ships are an important military and civilian target,and the study of ship detection in optical remote sensing images is a task with great potential in military and civilian fields.However,ship targets in optical remote sensing images have the characteristics of arbitrary orientation and complex background,which brings greater challenges to how to accurately locate ships.In addition,the number and scale of ships of various types are unbalanced.Both seriously affect the detection accuracy of the model.Aiming at the difficulties of ship detection tasks in optical remote sensing images,the main research contents of this thesis are as follows:1.This thesis designs a ship detection model SDFR based on the Faster RCNN model.SDFR introduces a rotation proposal region generation module and a rotation proposal region feature extraction module to obtain features with less redundant background,and uses a positive and negative sample selection module based on polygon IoU to define the positive and negative of training instances.Through these improvements,SDFR can accurately detect ship targets with arbitrary angles.2.This thesis further proposes an improved algorithm SDFR-AMF for the fine-grained ship detection task.In terms of network structure,SDFR-AMF improves the detection head structure,decouples the classification and regression branches in the structure,and uses the attention mechanism for feature enhancement,so that the detection accuracy of the model is improved.Aiming at the problem that the positive and negative samples of different scale instances are unbalanced by using a fixed IoU threshold in the training process,an adaptive positive and negative sample selection module is introduced to dynamically adjust the IoU threshold.Aiming at the problem of unbalanced number of samples,a data enhancement scheme based on ship instances is proposed.This scheme greatly alleviates the problem of low accuracy caused by unbalanced samples by cutting and pasting polygonal ship instances.3.This thesis conducts experiments on the public dataset HRSC2016.The results show that compared with the basic model Faster RCNN,SDFR and SDFR-AMF have improved the detection index mAP by 10.97 and 16.22 percentage points,respectively.In addition,detailed ablation experiments and hyperparameter sensitivity experiments are carried out in this paper,and the experimental results verify the effectiveness and robustness of the proposed algorithm.