Research On Target Detection Method Of Optical Remote Sensing Image Based On Deep Learning

With the rapid development of remote sensing technology,a large number of highresolution visible light remote sensing images have emerged.These images provide very important data support for the research in various fields of remote sensing.Among them,object detection plays a vital role in remote sensing image interpretation.Whether in the civilian or military fields,remote sensing image object detection has extremely high practical value.This paper mainly studies the optical remote sensing image object detection technology which is based on deep learning.Aiming at the characteristics of remote sensing images with many small targets,dense,and distributed in any direction,a detection algorithm suitable for remote sensing images is designed to improve the detection performance of remote sensing images.The content is as follows:A multi-scale remote sensing object detection method is proposed.Compared with natural images,remote sensing images have a more complex background,and the scene contains targets of multiple scales.Aiming at the problem of multi-scale target detection,this article is optimized on the basis of the Faster R-CNN detection framework.Optimization 1: Incorporate multi-scale module.First,we use a multi-path region generation network to generate regional candidate boxes of different scales at the multi-level feature map.Among them,the shallow feature map has a higher resolution,the perception of wildness is small,and is suitable for generating candidate frames for small targets;the high-level feature map has rich semantic information,the perception of wildness is large,and is suitable for generating candidate frames for large targets.Then,through the up-sampling operation,the deep and shallow features are stacked in the channel dimension to capture the multi-scale features of the target.Optimization2: Add attention mechanism.By introducing the operation of the attention mechanism,the network can autonomously learn a set of weights which representing the importance of each channel,enhance the features that play a significant role in the current task,and suppress the impact of complex background noise on the task.Optimization 3: Discussed the representation effect of the rotating rectangular frame.Because the targets in the remote sensing image are projected in any direction,when the targets are densely arranged,the horizontal rectangular frame cannot tightly wrap the target.Using the horizontal rectangular frame will usually cause excessive background redundant information,overlap of detection frames between targets,and missed detection.To this end,we introduced a rotating rectangular frame to refine the representation of the target,so that the prediction frame wraps the target more closely,and reduces the overlap between adjacent target prediction frames.We train and test on the HRSC2016 dataset and UCAS-AOD dataset respectively,and the results show that the optimized method has improved the detection accuracy of various scale targets,and the detection accuracy of the HRSC2016 dataset and UCAS-AOD dataset increased to 92.0% and92.2% respectively.A remote sensing object detection method based on center point is proposed.Since most object detection algorithms need to provide a priori frame to the target by setting an anchor frame,this method will bring a huge amount of calculation in subsequent Intersection over Union calculations and the non-maximum suppression process.In response to the above problems,this paper applies the Center Net algorithm to the remote sensing image object detection,and makes a series of optimizations on the basis of the algorithm.Optimization 1:Increase the angle information prediction.The angle information prediction is added to the branches of the original predicted target width and height to adapt to the characteristics of the target distribution in any direction.Optimization 2: Discuss a vector-based bounding box representation method.When returning to the width and height and rotation angle of the rotating rectangular box,the angle and the width and height are not in the same measurement unit,which makes the network difficult to train.Therefore,this article discusses a new bounding box representation.Optimization 3: Multi-task joint training.In order to improve the accuracy of object detection,the object detection and saliency detection task are jointly trained.And design related experiments on the HRSC2016 dataset and DOTA dataset for verification.Finally,the proposed method is compared with several current mainstream object detection methods,and the results show that this method has certain advantages.