Research On UAV Image Segmentation Based On Deep Learning

With the rapid development of UAV technology,a large number of UAV remote sensing platforms have appeared in recent years.The effective processing and key information acquisition of UAV images have become urgent problems in the field of remote sensing.UAV image segmentation,as an important part of image processing,plays an important role in promoting the development of remote sensing technology.The traditional segmentation method has a large amount of work and low accuracy;the segmentation method based on deep learning can extract the image features from the shallow to the deep and achieve the effective acquisition of semantic information.In view of the research status and existing problems,this paper takes the UAV image as the research object,and puts forward the corresponding improvement scheme from the aspects of image acquisition,quality improvement,neural network design,training and optimization.Firstly,aiming at the problem that the data set of semantic segmentation of high-resolution UAV image is scarce at present,the image is collected by Ruiying rs100 e UAV equipment,and the improved overlapping region enhancement fusion algorithm is used to realize the fusion of surface image,and a high-resolution UAV surface image data set is made.There are eight different landforms in the data set,which can be used as the benchmark to evaluate the performance of semantic segmentation model and as the material for other remote sensing image interpretation tasks.Secondly,aiming at the problem of poor image quality in the process of UAV image data acquisition,the image data is preprocessed from two aspects of edge enhancement and image denoising.The improved unsharp mask method is used to enhance the edge details of the image,and the improved method based on wavelet threshold denoising is combined to reduce the noise and improve the quality of UAV image.Then,aiming at the problems of low accuracy of UAV image segmentation algorithm in complex scenes and imbalanced data sets(long tail data),a dam deeplab model is proposed based on deeplab network and optimized from data evaluation and model structure.The specific optimization process includes replacing the original backbone network;In order to solve the problem of insufficient data,the training parameters of backbone network on Imagenet model are retained,and the network training parameters are shared through transfer learning;Attention mechanism module is added to guide feature learning;According to the characteristics of UAV image,the expansion rate combination of expansion convolution space pyramid pooling module is adopted;To solve the long tail problem in data sets,the cross entropy loss function is replaced by the class balance loss function.Finally,the model is trained and tested on the UAV image data set.The experimental results show that the proposed dam deeplab model achieves 77.3% of the average PAPR and87.7% of the average pixel accuracy on the test set.Compared with the original model,the image segmentation effect is significantly improved.