Research On Remote Sensing Image Object Detection Based On Deep Learning

With the rapid development of high-resolution satellite and UAV technology and the advancement of imaging technology,the scale and quality of aerial optical remote sensing image data have been greatly improved.Being able to quickly and accurately process the data of massive high-resolution remote sensing image is crucial in current research focusing on remote sensing image object detection.Currently,though significant progress has been achieved by the object detection algorithm based on deep learning on natural images.However,there are obvious differences between remote sensing images and natural images,such as large scale changes between targets of the same category,large number and dense distribution of target objects,etc.Therefore,directly applying the existing deep learning object detection algorithm to the field of remote sensing technology is an urgent need to solve.In order to better achieve the remote sensing image object detection,this paper has done the following research work:(1)An improved SSD algorithm is proposed.Learn from the idea of fusing high and low semantic information with Feature Pyramid Network(FPN),enhance the semantic information of shallow feature maps and the position information of deep feature maps,improve the method of multi-scale feature fusion of SSD algorithm,use the generalized Intersection over Union(GIoU)loss function to weigh the distance between the predicted box and the ground true box,and solve the problem that the Intersection over Union(IoU)loss function cannot be optimized when the two boxes do not overlap.The experimental results on the public remote sensing image data set DIOR show that the improved SSD algorithm in this paper improves the mean Average Precision(mAP@0.5)of the original SSD algorithm by 2.5%to 61.1%.(2)Aiming at the problem of poor detection of small-scale targets in remote sensing images,a remote sensing image object detection model based on three different depths of YOLOv5 is constructed.The YOLOv5 algorithm uses a structure similar to a Path Aggregation Network(PANet),which shortens the path for the propagation of shallow and deep feature information,and makes the better use of the detailed location information in the shallow features.The backbone network has also designed multiple cross stage partial structures to ensure accuracy while reducing the weight of the model.The mAP@0.5 of the YOLOv5s,YOLOv5m and YOLOv51 models on the DIOR dataset are 65.4%,69.8%and 71.2%,respectively.(3)A method to improve the YOLOv5 model is proposed.The channel attention mechanism layer is introduced into the backbone network to strengthen the expression of the correlation between the target information of the feature maps and the important channels.The FReLU activation function is used to improve the network's analysis of irregular target objects.Ability to further improve the accuracy of object detection in optical remote sensing images.Through multiple experimental comparisons,the improved YOLOv5s-SE-FReLU model in this paper is 1.8%and 2.4%higher than the mAP@0.5 and mAP@0.5:0.95 of the YOLOv5s model,respectively.(4)Simplified the remote sensing image object detection model based on YOLOv5,completed the model deployment of the Android mobile terminal by using the NCNN framework,and achieved the whole process from theory to model design to practical application.