Research And Application Of Object Detection Method In Remote Sensing Image Based On Lightweight Network

In recent years,with the aviation industry and remote sensing technology developed by leaps and bounds,the remote sensing images collection are getting easier.There many large remote sensing data sets have emered.The resolution of the remote sensing images is getting hiher,and the images contain a wealth of information.Which is great help to the application of remote sensing images such as agricultural monitoring,military reconnaissance,autonomous driving,etc.Mean while,exreacting information from the remote sensing images becomes more and more difficulty,because of the images are diversity and complexity.Target detection plays an extremely important role in the field of remote sensing images.With the rapid progress of computer vision and deep leaning,the target detection alforithms have sprung up.These algorithms have powerful learning capabilities and generalization capabilities.Among them,the regression-based algorithms represented by You Only Look Once(YOLO)show a good balance between accuracy and speed.Therefore,this thesis utilizes the YOLOv3 algorithm as the basic framework for target detection in remote sensing images.However,the computing equipment have problems such as the limited of the ability to calculate and the small memory.We introduce the ideas of design lightweight network to enable the overall algorithm to run in real time.The main contents as follows:1.Target detection in remote sensing image based on lightweight YOLOv3.First,we designed a lightweight feature extraction by utilized the depthwise separable convolution,which is a convolutional decomposition method,and can greatly reduce the parameters.Secondly,we embed the attention module SE in the network,which can further enhance the description of feature capabilities by selecting enhancement information and suppressing useless information.Finally,a spatial pyramid pooling module is added to the neck of the network,which can ensure translation invariance by aggregating local features.It also can filter more irrelevant pixel values,and help to get better recognition features for classifier.Experiments reveal that the propsoed algorithm can maintain high efficiency and detection accuracy simultaneously.2.Target detection in remote sensing image based on rotatable bounding box.Aiming at the problems of the target with orientation and scale variations in the remote sensing image scene.The angle penalty is added to the method proposed in Chapter 3.The angle loss is redesigned based on the GIo U loss function.Moreover,adjusted and optimized the structure of YOLO.The orientation response convolution can rotate actively,which can help to match the rotatable bounding box.Therefore,the algorithm can detect with rotation angle.Experiments show that the algorithm has advantages in accuracy and speed.The accuracy of rotating targets detection on the large-scale remote sensing image data set(DOTA)can reach good effect.