Research On Remote Sensing Images Object Detection Method Based On Deep Learning

As the resolution of remote sensing images increases,the information obtained from remote sensing images becomes more and more abundant.Remote sensing image recognition technology is of great significance in military navigation and environmental monitoring.It has important practical significance and application prospects for object detection on remote sensing images.However,there are many interference factors in remote sensing images,such as special angle of view,high background complexity and small target problems,etc,which makes higher requirements and challenges for remote sensing images object detection technology.In recent years,deep learning has made remarkable achievements in the field of image classification and recognition,which has attracted the attention of many researchers.Although deep learning methods have significant advantages in terms of detection performance over traditional methods,due to the particularity of remote sensing image,the current remote sensing image object detection algorithm can not meet the requirements of large-scale automation applications in terms of detection rate and accuracy.The detection rate and detection accuracy can be greatly improved.In view of the above problems,the main work and innovations of this paper are as follows:(1)An improved SSD(Single Shot MultiBox Detector)method for remote sensing images object detection is proposed.Firstly,aiming at the problem that the parameters of VGG-16 are various and the calculation cost is high,inception-v2 is proposed as the basic network in SSD model to reduce the time spent in feature extraction and accelerate network training.Secondly,aiming at the problem that small object detection is not effective in multi-scale object detection in SSD,the Mixed 3c layer,Mixed 4c layer and Mixed 5c layer in inception-v2 are connected and fused by upsampling and Concat operation,so that the low-level feature semantic information can be combined with the high-level feature semantics,and the ability of extracting features from the low-level network can be improved.Finally,Focal Loss is proposed to solve the problem of imbalance between positive and negative samples in SSD training process,which makes SSD network focus on hard-to-classify samples and accelerate network convergence.Comparative experiments were carried out on three remote sensing images datasets: NWPU-VHR,UCAS-AOD and RSOD.Compared with the original SSD,the mean average precision was increased by 5.4%,6.1%,and11.6%,respectively,while the speed was matched.(2)An improved Faster R-CNN method for remote sensing images object detection is proposed.In view of the large proportion of small and medium objects in remote sensing images,through the detection performance analysis of Faster R-CNN,firstly,it is proposed to add Conv2 2 layer,Conv3 3 layer and Conv4 3 layer as input of RoI pooling layer in Faster R-CNN network structure.The input of the RoI pooling layer is increased from the previous Conv5 3 layer to 4 layers,and the output of RoI pooling layer is fused by Concat operation as the final feature map for prediction.Secondly,matching the scale of the objects in the remote sensing images by changing the scale of anchors in Faster R-CNN,the number of anchors changed from 9 to 15.Compared with the original Faster R-CNN,the mean average precision on NWPU-VHR,UCAS-AOD and RSOD datasets was increased by 2.5%,2.7%,and 3.6%,respectively.