Research On Remote Sensing Image Target Detection Method Based On Deep Convolutional Neural Network

The target detection of remote sensing images is an important basis for image interpretation tasks.With the improvement of the resolution of remote sensing images,the target detection of high-resolution remote sensing images has become a hot issue in the field of remote sensing image processing.Compared with the feature extraction method designed by general manual,the target detection algorithm based on deep learning has good robustness and detection.The accuracy of the results is high,and a key breakthrough has been made in the detection of natural images.However,the target detection algorithm based on natural images is directly used in the remote sensing image target detection,and high-precision detection results cannot be obtained.This is because the high-resolution remote sensing image target itself has variable shapes,different sizes,and a complex background and dense distribution.It brings new challenges to the detection task.Therefore,this paper combines deep learning theoretical knowledge to study the remote sensing target detection algorithm based on convolutional neural network.The main research work of the paper is as follows:(1)An algorithm based on YOLOv3 deep learning target detection is presented.It improves the basic feature extraction network and the training loss function,and trains and generates a model that is beneficial to remote sensing feature detection.According to the constructed remote sensing image target detection data set,a new network structure Darknet46 is designed,which can more effectively extract the depth features of the image and improve the detection rate of the network;Design a suitable clustering center;improve the loss function of YOLOv3,and prove the effectiveness through experiments.Finally,comparison experiments with typical algorithms YOLOv3 and Faster R-CNN show that the improved model can obtain higher mAP for three different types of ground objects,which improves the accuracy and recall rate to a certain extent.(2)A target detection algorithm based on Faster-RCNN for high-resolution remote sensing images is proposed.Replacing VGG16 with the network structure of residual convolutional network ResNet50 and feature pyramid structure,and replacing the conventional convolution method with hollow convolution to expand the receptive field and realize the effective extraction of deep features of complex features;adopt ROI Align replaces the ROI Pooling operation to solve the regional position deviation caused by quantization in the ROI Pooling operation;replace the traditional Non-Maximum Suppression(NMS)method with the improved Soft Non-Maximum Suppression(Soft-NMS)method Post-processing improves the error suppression and easy overlap of the detection frame of the NMS algorithm.Experimental results show that compared with YOLOv3 and Faster R-CNN,the improved Faster R-CNN algorithm can adapt to the characteristics of large scale changes in remote sensing targets on the public DIOR data set,and the detection effect on small targets is better,and the accuracy is remarkable.