Research On Improvement Of Super-Resolution Color Remote Sensing Image Reconstruction And Target Detection Algorithm

The The rich semantic perception information,high-frequency detail information and special imaging perspective make remote sensing data and images widely used,among which the panchromatic remote sensing images,though with high spatial resolution,lack spectral information to show the color of the feature is not conducive to target detection;while the target detection of color remote sensing images gradually becomes a research hotspot,but the problems are:remote sensing targets are densely distributed,occupy a small size and have a complex background.Although the target detection of high-resolution color remote sensing images is more advantageous,it is more difficult to upgrade the hardware equipment due to the high accuracy requirements in the acquisition process;therefore,this research work focuses on the following three aspects.1.In order to obtain color remote sensing reconstructed images containing more high-frequency perceptual information and texture detail information,a super-resolution reconstruction algorithm for remote sensing images with a generative adversarial network that incorporates multi-scale perceptual field modules is proposed.Firstly,the global feature acquisition is enhanced by using multi-scale convolutional cascade to remove the normalization layer in the generative adversarial network to improve the network training efficiency and remove the artifacts and reduce the computational complexity;secondly,the multiscale perceptual field module and dense residual module are used as the detail feature extraction of the generative network to improve the network reconstruction quality and obtain more detailed texture information;finally,the Charbonnier loss function is combined Finally,the Charbonnier loss function and full variational loss function are combined to improve the training stability of the network and accelerate the convergence.After the experimental study,the improved algorithm of super-resolution remote sensing image reconstruction based on generative adversarial network proposed in this paper has a higher peak signal-to-noise ratio of about 1.65 dB,higher structural similarity of about 0.040(5.2%)and higher feature similarity of about 0.010(1.1%)than that of super-resolution generative adversarial network.2.Based on the YOLOv4 target detection algorithm,the Focus structure is introduced into the backbone feature extraction network for the problems of dense and small-scale targets and complex and unevenly distributed backgrounds in remote sensing images,retaining more semantic information of image initialization,fusing multi-layer separable convolution,and improving the backbone feature extraction capability by sharing feature maps in the same dimension with different scales of perception without deepening the network.network feature extraction capability.To address the problem of multi-scale target detection in remote sensing images,a bidirectional feature pyramid network is used for multi-scale feature fusion,with multiple top-down and bottom-up network structures,integrating shallow high-frequency detail information with deep global information,introducing lightweight sub-channel attention mechanism,and using deep separable convolution to enhance multi-scale feature integration capability.After the experimental study,the improved remote sensing target detection algorithm based on YOLOv4 proposed in this paper has improved the average accuracy by 16.17%on DIOR dataset and 2.87%on RSOD dataset compared with YOLOv4 algorithm.3.The original images of DIOR dataset were used as real scenes and downsampled at different multiples of resolution,and the detection accuracy decreased significantly when the image resolution was lower than half of the input pixel threshold of the detection algorithm;when it was lower than one-third of the threshold,the detection accuracy decreased by more than 50%.The low-resolution NWPURESISC45 dataset was reconstructed with super-resolution remote sensing images,and the average accuracy of the reconstructed images was improved by 29.3%compared with that of the original image detection.In this paper,we use neural network algorithm to enhance the resolution of remote sensing images and improve the target detection algorithm of color remote sensing images to solve the problems of difficult training of super-resolution reconstruction algorithm and missing details of reconstructed images,so as to achieve high accuracy detection of medium and low resolution color remote sensing targets.