Research On Salient Object Detection Of Hyperspectral Image Based On Deep Network

With the rapid development of hyperspectral remote sensing and data analysis technology,the object detection of hyperspectral image obtained by imaging spectrometer has gradually become one of the important research directions in the discipline.Especially salient object detection,which belongs to a method of object detection combined with human visual mechanism.Salient object detection can make the object in the scene stand out from the background.And no manual intervention is required.Therefore,it’s practical.The traditional salient object detection algorithm generally adopts the method of direct dimensionality reduction combined with differential calculation,which often loses image information and does not extract enough features for salient objects.So how to make full use of the spatial spectrum information of hyperspectral data,and realize the rapid location of salient areas in complex scenes,is an urgent problem to be solved in the current application.As one of the representative algorithms of deep learning,convolution neural network(CNN)can extract general high-dimensional features from hyperspectral data which are very helpful for data analysis.Combining with the technology of computer vision and starting from the characteristics of hyperspectral image,this paper proposes a salient object detection algorithm combined with deep network,aiming at the shortcomings of traditional algorithm.Then we use the open and real hyperspectral data for experiments.The main work of the paper includes:1.The development status of hyperspectral remote sensing technology is studied,and the detailed concepts of object detection and salient object detection of hyperspectral image are introduced.2.The data format and data acquisition method of hyperspectral image are briefly described;the basic information of the open datasets for salient detection is given;We constructed the experiments and collected the image by SWIR-384 spectrometer,which is used for subsequent experiment and comparison.Meanwhile,the traditional algorithm and evaluation index of salient object detection are introduced in detail.3.To solve the problems of traditional detection algorithm,this paper proposes a detection algorithm based on weakly supervised deep learning(SWDL).The algorithm attempts to extract the deep features of salient objects by using CNN,and solves the problem of manual labeling information needed for training the network.The algorithm combines the method of low rank matrix recovery,and extracts the spatial spectrum characteristics of pixels through three-dimensional CNN,then,defines the significance of pixels.The results show that the average AUC value and F-measure value of the proposed algorithm on HS-SOD data set reach 85% and 88% respectively,which are better than other comparative algorithms and have better detection effect.4.Aiming at the slow speed of SWDL algorithm,a algorithm based on unsupervised segmentation network(SDUN)is proposed combined with unsupervised image segmentation network.The ability of SWDL algorithm to extract features through CNN is excellent,but the training process still consumes lots of time.SDUN algorithm use unsupervised network for a process of dimensionality reduction and feature extraction by 1D convolution and 2D convolution.Then,the salient results of image are obtained by using manifold sorting technology.This process is a self-iterative process without network training.The experimental results show that the average AUC value and F-measure value on HS-SOD data set reach 89% and 94%,respectively.The algorithm has good detection effect and can greatly reduce the calculation time.This paper studies the problem of salient object detection of hyperspectral image,analyzes the shortcomings of traditional algorithms,proposes two detection algorithms combined with deep network,and carries out experiments on HS-SOD and SDA.The experimental results show that the detection effect of the proposed algorithm on HSSOD and SDA has different degrees of improvement compared with the traditional algorithm.For the actual application of the detection effect and the speed,the two algorithms have their own advantages.