Research On Target Detection Algorithm Based On Remote Sensing Airplane Image

Under the current situation,as special military equipment,it is very important to accurately detect and classify the aircraft.Among them,the remote sensing aircraft image is a difficult point to detect.Because it has more complex background information than ordinary images,the target direction is chaotic and small targets are mostly.Because the shooting process is greatly affected by weather factors,most of the images are blurred.These characteristics lead to many target detection algorithms applied on conventional images,which show unsatisfactory results when used to detect remote sensing images.The overall detection accuracy is low,and there is a certain phenomenon of miss detection.When the target to be detected is small and fuzzy,the detection effect will be worse.To solve these problems,a new method for remote sensing aircraft image detection is presented.Based on the deep learning method,this paper detects and classifies remote sensing aircraft images,and proposes a new target detection network backbone.The overall network adopts the encoder-decoder architecture and the idea of global and local multi-scale feature fusion,so that the extracted target sample features are more sufficient and accurate,to achieve more accurate detection and classification of remote sensing aircraft images.The specific work is as follows:For a small target in remote sensing image and insufficient feature extraction of the fuzzy target,a global and local multi-scale feature fusion target detection network GLF-Net based on encoder-decoder architecture is presented.The overall network consists of preprocessing module for image preprocessing,a feature extraction module for extracting multiscale features of target samples,and a classification regression module.The first step is to send the measured image to the preprocessing module to perform some basic operations on the image,such as uniform size.Then,the feature extraction module is input for feature extraction,in which the image is first extracted from the encoder part,then the feature information is sent to the improved attention mechanism module for processing,then into the decoder part for global feature extraction,and then the feature information extracted from the two parts is fused into the classification regression module.The final output of the test results.The test results in the self-made dataset show that the network has a more accurate detection accuracy,and subsequent generalization performance experiments also show that it has a better generalization performance..Aiming at the problem that the detection effect of GLF-Net on difficult samples is not ideal,considering that the problem is caused by insufficient global feature extraction,we improve the decoder part of the network based on GLF-Net,further expand the receptive field,aggregate the network context information,and enhance the ability of aircraft global feature extraction.The network model is named DP-Net.To prove the detection ability of DP-Net on difficult samples,a difficult sample detection data set is established to evaluate the performance of GLF-Net and DP-Net.The experimental results show that DP-Net improves the poor detection performance of GLF-Net for difficult samples.