Design And Implementation Of Object Detection Algorithm Based On Deep Learning In Remote Sensing Images

In recent years,remote sensing image recognition has gradually become a basic but challenging research topic in the field of aerospace and satellite image analysis.The deep learning algorithms,which learn the high level discriminative features from large-scale data,have become a hot topic in the field of computer vision and have been gradually introduced into the field of remote sensing image recognition.In this thesis,we design an efficient object detection method to deal with multi object detection problem in remote sensing images.Unlike ordinary natural images,the scene in remote sensing images is more complex,the image quality is poor and there are a lot of small scale targets such as vehicles,ships etc.In order to ensure the accuracy of the detection of various scale targets,this thesis proposes a new object detection method called Scale Adaptive Convolution Neural Network(SA-CNN),which applies the Faster R-CNN detection framework based on Convolution Neural Network(CNN)to detect objects.First,the Dual-path Region Proposal Network(DRPN)uses the multi feature layers to produce multiscale object proposals.The lower-level features are better for small target proposal generation with the higher resolution and smaller receptive field,and the higher-level features are more suitable to produce large region proposals for the stronger semantic and wider receptive field.Furthermore,in order to enrich the information in features,we merge the low-level features with the high level features by adding them up.By this way,it can improve the final prediction of network without increasing two much parameters.The training and testing of the remote sensing image dataset NWPU VHR-10 shows that SA-CNN not only guarantees the detection accuracy of the large target objects,but also improves the accuracy of the detection of small targets.The average accuracy of all categories is increased to 89.2%,and the average detection time of single remote sensing image is 0.336 s test on GPU NVIDIA GTX 1080 ti,which shows the efficiency and stability of SA-CNN.