Hyperspectral Image Object Detection And Scene Classification Based On Deep Learning

Compared with common images,hyperspectral images have huge advantages in many aspects.As the cost of acquiring hyperspectral images is getting lower and higher and the accuracy is getting higher,researching how to use hyperspectral images to better serve our life become more and more important.In this thesis,based on the problems in real life and existing algorithms,combined with the advantages of hyperspectral images,a new task and corresponding implementation algorithm for object detection and scene classification based on hyperspectral image data are proposed.The specific research contents are as follows:1)Since there is no public dataset available for the regional garbage detection task,it cannot meet the training needs of the general object detection model.Based on the characteristics of single-pixel classification based on hyperspectral images,the use of hyperspectral image classification algorithms to generate binary segmentation maps of regional garbage reduces the difficulty of target detection so that it can be directly detected with unsupervised object detection algorithms and achieve a high detection accuracy,in order to validate ouralgorithm,we have labeled a hyperspectral image object detection dataset for detecting regional garbage.2)In view of the problem that the classification accuracy of common image scenes is not high in some scenes,this thesis proposes the scene classification task of hyperspectral images.Scene classification based on hyperspectral images can use the rich spectral information of hyperspectral images to improve the classification accuracy.In addition,for the problem of large background differences in remote sensing images,a hyperspectral scene classification model based on global and important region features is proposed.This model can filter out complex background interference without increasing the model parameters.3)The characteristics of large data and large parameters of the convolutional neural network model limit its application in tasks with limited computing resources and data resources.This thesis proposes a hyperspectral image scene classification model based on graph convolutional neural network.This model determines nodes and node features through clustering module,then creates graphs with preset graph edges,and extracts features through graph convolution layers.Finally,fully connected layer is used to output the classification results.The model successfully achieved satisfactory accuracy under the premise of using very few parameters.