Research On Aerial Image Resolution Method Based On Deep Learning

The scene parsing of aerial image based on deep learning is a research focus in applications of UAV Technology and deep learning.At the same time,some technologies have been widely used in drone investigations,traffic supervision,land surveys,wildlife tracking,and disaster observation.Aerial images are easily affected by bad weather,resulting in blurred images,reduced contrast,color fidelity,and sharpness,therefore,it is necessary to perform defusing aerial images and enhancing images to provide high quality for target detection.The existing target detection network models based on deep learning usually use images with a resolution of 300*300 to 500*500,and can achieve real-time detection,and the detection accuracy is far superior to the target detection method based on machine learning.Directly compressing the high-resolution aerial image will lose many details of the small target,and the resolution of the feature map in the network model is low,and there is little detailed information about the small target retained,which is not conducive to small target detection.At present,the image defusing methods and target detection methods based on deep learning does not apply to high-resolution aerial images,there are some problems,such as image defusing algorithm can't handle the sky area aerial image for high resolution Aerial image,the target detection network model has poor positioning accuracy for small targets,and the detection time does not meet the real-time requirements.In this paper,the thesis introduces the basic theory of image defusing algorithm based on dark channel and the classical deep convolution neural network model,the aerial image defusing algorithm based on guided filters and two kinds of aerial image target detection based on deep learning are mainly studied.The proposed target detection methods are target detection method based on a pre-segmented deep learning and a feature-based deep learning aerial image target detection method.The image defusing algorithm and two kinds of aerial image target detection methods based on deep learning are verified by experiments and compared with the results of the existing classical algorithms.The main work and contributions of the paper are as follows:1.A fast defusing algorithm for aerial images based on guided filter is proposed in this paper.Firstly,the atmospheric screen function of aerial image is calculated by the dark channel,and the edge preservation estimation of atmospheric screen function is calculated by the guide filter,which avoids the calculation of the atmospheric light value and transmittance map from the dark channel;then the residual image of the original image and the atmospheric curtain function is calculated.And extract its light channel;Finally,according to the light reflection model to obtain the reflection coefficient of the scene as the restoration image.Compared with the classic single image defogging algorithm based on dark channel,this method can obtain better image detail and overall visibility without global atmospheric light,which is more suitable for aerial images.2.The aerial image target detection method based on pre-segmented and deep learning is proposed in this paper.At present,the candidate region extraction algorithm in the existing network model takes too long to meet the real-time requirements.Therefore,a fast candidate region extraction method based on quadtree segmentation is proposed,and then features and target classification and location are extracted through a CNN,which can improve the detection accuracy and ensure the target detection network real-time.We propose a target detection method based on feature fusion and deep learning in aerial images.The last low-resolution feature maps of the traditional deep learning network can predict target,which only have high-level semantic information and lack shallow information about the target location,so the positioning of small targets is inaccurate.Aiming at these problems,a deep learning target detection method based on feature fusion is proposed.By adding a feature pyramid network and adding transverse connections between convolution layers in the network,the low-resolution high-level semantic features are achieved through upsampling.High-resolution shallow edge features are combined to improve the positioning accuracy of small objects.