Research On Object Detection In Remote Sensing Images Based On Dynamic Lightweight Convolutional Neural Network

Optical remote sensing target detection technology has been widely used in many industries and fields.In the investigation and protection of natural resources,optical remote sensing target detection is helpful to effectively grasp the transformation of key resources.In urban planning,optical remote sensing target detection has important applications.In military application,optical remote sensing image target detection can help to carry out accurate target attack.In recent years,with the rapid development of convolutional neural networks,object detection algorithms in optical remote sensing images have been widely used.At present,object detection algorithms based on deep learning can be divided into two categories: one is a single-stage convolutional neural network object detection algorithm represented by YOLO(You Only Look Once)series,and the other is a two-stage convolutional neural network object detection algorithm represented by R-CNN(Region-CNN).Considering that convolutional neural network model has many parameters and a large amount of computation,it is difficult to deploy in various edge computing situations.Therefore,lightweight convolutional neural network has become an important research direction,which has the characteristics of fewer network parameters and less computation,and can meet the operation requirements of embedded devices.The dynamic convolutional neural network can dynamically adjust the network structure and network parameters,and adaptively construct the convolutional neural network structure and select the optimal internal network structure parameters.In this paper,lightweight and dynamic network construction methods are combined to achieve high-resolution optical remote sensing target detection based on dynamic lightweight convolutional neural network.The proposed method has many advantages,such as fewer model parameters,fast training and online detection,model performance and detection task complexity,and dynamic balance between computing resources.The following is the primary research content:1.A lightweight convolutional neural network method is suggested in this paper.This approach uses the Squeeze Net lightweight backbone extraction network in place of the YOLOv3 object detection algorithm’s original feature extraction network.The YOLOSqueeze Net lightweight convolutional neural network is implemented,and experiments are conducted on our own wide area key building object detection dataset.Experiments show that the YOLO-Squeeze Net lightweight convolutional neural network used in this paper reduces the size of network parameters and improves the detection speed while ensuring the accuracy.2.In this paper,a dynamic lightweight convolutional neural network remote sensing target detection method is proposed.In the remote sensing image target detection task,the dynamic network scalability,combined with the adaptive selection of data sets,and the lightweight detection model proposed in this paper are used to achieve the matching of target detection type accuracy and model structure.Experimental results show that the Dark Net-53 dynamic convolutional neural network m AP proposed in this paper reaches 92.08%,which is higher than many mainstream target detection models,reduces model parameters and effectively avoids model overfitting.The lightweight dynamic convolutional neural network m AP of Squeeze Net reached 88.70%.Compared with the mainstream benchmark target detection model,based on the above advantages of dynamic detection network,only a small amount of precision was sacrificed to further greatly reduce network parameters and achieve fast training and online detection.