Research On Aircraft Target Detection Algorithm For Color Remote Sensing Images Based On Deep Learning

The color remote sensing image has a large amount of information and contains important perceptual information.It has important application value for the automatic detection of specific targets of interest.In recent years,deep learning has performed well in object detection of natural images due to its strong nonlinear processing capabilities.However,due to the target attributes of remote sensing images and natural images,directly applying them to remote sensing images will cause a lot of problems,especially for the attributes of aircraft targets of different sizes,small targets,and multiple attitudes in remote sensing images,which make more efficient target detection algorithms are needed.Facing the needs of aircraft target detection in color remote sensing images,this paper is based on deep learning to conduct some research on aircraft target detection in color remote sensing images.(1)The edge chromaticity of color remote sensing images has large differences,and rich semantic information can be obtained for edge detection.Therefore,an edge detection algorithm for color remote sensing images is proposed:Using a combination of fractional differential and Canny operators,a fractional differential template is designed to be applied to Canny operators.At the same time,the Gaussian curvature filtering theory is introduced to smooth the image while performing edge detection on color remote sensing images to find the smallest regular The energy points are optimized to obtain the optimal parameters of the algorithm,so as to obtain a better edge detection effect of the color remote sensing image.Experiments show that compared with other four common edge detection algorithms,this algorithm can effectively suppress the non-linear amplification and diffusion of noise during edge extraction of remote sensing images,and retains rich image texture information.(2)This paper does target detection for aircrafts in remote sensing images,but the data set lacks resources.In order to obtain better training results,a remote sensing image aircraft target detection dataset is constructed.Through cropping,rotation,denoising and improved edge detection algorithm processing,1000 color remote sensing images containing aircraft targets were obtained.Then,manually label according to the format of PASCAL VOC data set,and divide it into training set,test set and validation set through the program.Through data set comparison experiments,it is found that the data set constructed in this paper can obtain more information during network training,which helps improve the accuracy of the target detection algorithm.(3)Aiming at the problems of missed detection and false detection in the target detection process caused by the small,multi-size and different attitude of aircraft targets in color remote sensing images,a color remote sensing image aircraft based on Faster R-CNN and FPN network Objective improvement model.Focus on analyzing the RPN network and FPN network in Faster R-CNN,combining the advantages of the two networks to improve the model.Without additional calculation,the FPN network is embedded into the feature extraction module of the Faster R-CNN network to improve the feature extraction method.Through bottom-up paths,top-down paths,and horizontal connections,low-resolution,strong-semantic features are combined with high-resolution,weak-semantic features to form feature map sets,which are output at each level.After training experiments on aircraft target datasets of remote sensing images constructed by ourselves,it can be found that the addition of the FPN network enables the entire target detection model to obtain richer feature information,ensuring that aircraft targets can be correctly identified during classification detection and positioning An accurate bounding box is given to realize the detection of aircraft targets.Experimental research proves that the detection algorithm for aircraft targets based on deep-learning color remote sensing images proposed in this paper is 18.6%higher than the SSD,R-FCN and faster R-CNN models in detecting aircraft target incompleteness and target blurring.