Research On Optical Remote Sensing Object Detection Technology Based On Deep Feature Enhancement

Optical remote sensing image object detection is widely used in various fields of civil and military applications,and is a very important research direction in the field of computer vision and pattern recognition research.Feature extraction is a key step in object detection.Deep learning can automatically extract features,make full use of big data,and end-to-end training and prediction.However,existing depth detection models lack the cognition of model features,and the extracted feature information is single.In this paper,the object detection algorithm of deep learning feature enhancement is studied.Three remote sensing image object detection methods based on depth feature enhancement are designed,which realizes efficient detection and recognition of optical remote sensing image targets.The main research contents and results of this paper are as follows:1.Aiming at the multi-scale characteristics of remote sensing image targets and the problem that the target is susceptible to background similar objects,a dual path channel attention mechanism detection model DPCA-Det is proposed.Introducing a dual-path channel attention module in the classification prediction layer of the detection model,wherein a path attention module learns the importance of each convolution channel of the classification prediction layer,and weights the importance of each channel per channel to the classification prediction layer.The feature regions associated with different convolution channels are different to increase the characteristic response of the target channel,and the characteristic response of the noise region channel is weakened;another path attention module extracts the key features of each channel of the classification prediction layer,and directly connects through the short connection Acts on the classification and output layer of the model to enhance the features of the classification output.Experiments on the UCAS-AOD object detection data set show that the method is compared to Yolov3,SSD-512,Retina Net,Faster-RCNN,R-FCN,R-FCN-OHEM,FPN,Refine Det models,in UCAS-AOD.The data set m AP increased by 1.4%-7.6%,and the NWPUVHR₁0 data set increased by 2.6-8.6%.2.According to the DPCA-Det model proposed in the previous chapter,the accuracy of small object detection is not high.The small target detection model Fine-DPCA-Det with adaptive negative sample learning is proposed.The deconvolution module is introduced to improve the prediction layer characteristics.The resolution,design small-scale anchor frame and introduce context information to guide the detection of small targets,and introduce the cavity convolution to increase the feature receptive field.In order to further improve the discriminability of features,an adaptive negative difficulty learning method is designed.Validation on the UCAS-AOD dataset shows that the proposed method improves m AP compared to Yolov3,SSD-512,Retina Net,Faster-RCNN,R-FCN,R-FCN-OHEM,FPN,Refine Det,and DPCA-Det models.About 0.7-8.3%,in the NWPUVHR₁0 data set and DPCA-Det equivalent,the m AP in the IPIU test data set increased by about 0.3%-2.3%.3.Aiming at the shortcomings of wide-speed remote sensing image overlapping sliding window detection method with high speed and high false detection rate,a wide-range remote sensing object detection method based on region of interest re-detection is proposed.Firstly,the tower sample set is made centered on the target during the training process.In the test process,the potential target is extracted by the fast sliding window without overlapping,and the new image area is sampled in the wide image for re-examination with the potential target as the center.That is,the input of the network model is adjusted and enhanced,so that the test input fits the distribution of the training data.In the IPIU wide-range remote sensing image test dataset,using SSD as the detection model,this paper proposes that the region of interest re-detection method is about 1.8%-3.4% higher than the sliding window overlap detection method,and the detection time is reduced by up to 115 s,using YOLOv2 The detection model,this paper proposes that the region of interest re-detection method is about 2.1%-3.8% higher than the sliding window overlap detection method,and the detection time is reduced by 64 s at most;the interest region re-detection method has higher detection accuracy than the sliding window overlap detection method.faster.