A Study On The Target Detection Algorithm Of High-resolution Hyperspectral Images

The target detection of hyperspectral images aim at pointing out all the positions of targets from hyperspectral images,which is a technique about spectrum analysis and computer vision.It is a branch of hyperspectral images processing,which is widely applied in the fields of military,agriculture,mining,etc.With the development of the unmanned aerial vehicle(UAV)technology and the spectral imaging technology,the high-resolution hyperspectral image(HRHSI)captured from the UAV was born at the right moment.Compared with the traditional hyperspectral image which has a low resolution,the HRHSI has its own characteristics,such as high resolution and big data.It is difficult to achieve a satisfied result when traditional target detections directly apply to the HRHSI.Therefore,it is urgent that a new target detection need to be developed for HRHSI.This paper proposes an effective target detection algorithm for the HRHSI.A number of studies have been carried out.The main researches are introduced as following.(1)This paper proposes a novel corner feature for HRHSI named extreme-constrained spatial-spectral corner(ECSSC for short),and the extraction algorithm of ECSSC.Since the corner feature plays well in image processing and pattern recognition,and the extraction of corner from the three-dimension data is able to fully consider the spatial and spectral signatures of the hyperspectral image,this paper proposes a new corner feature for hyperspectral image and its extraction algorithm.We extract the feature points where the data values change rapidly in spatial domain and smoothly in spectral domain.The experimental results show that the proposed algorithm can detect abundant corner features with high repeatability rate from HRHSI and the accuracy of image-level HRHSI based on ECSSC is dramatically higher than that of the existing classification algorithms.(2)Based on ECSSC,this paper proposes a novel target detection for HRHSI.The algorithm includes two stages.In Stage one,based on the proposed ECSSC and the proposed spatial pyramid matching model called spectral curve spatial pyramid matching(SC-SPM)model,we are able to search the potential target regions.In Stage two,a novel bounding box dilation is proposed to detect the target region precisely by utilizing the SC-SPM kernel and image representation.The results demonstrate the superiority of the proposed algorithm compared to those traditional target detection methods in different situations(e.g.different scenes,illumination and scales).(3)This paper proposes a novel target detection for HRHSI based on the two-stream feature-extraction network.Since the targets are generally rare in HRHSI,and the loss of the neural network is hard to be convergent wihout a large quantity of data,this paper proposes a novel target detection for HRHSI based on the two-stream feature-extraction network.One stream is in the spatial domain for the feature extraction by the convolutional neural networks(CNN),the other one is in spectral domain.Then,the features that extracted by two streams are fused.Finally,the training data are utilized to construct a target-background feature space(T-B feature space).In this space,the target can be easily detected by calculating the target confidence of the unknown samples.The results indicate the proposed algorithm is better than the others.(4)This paper designs a target detection system of HRHSI.This system includes the hyperspectral images visualization,band synthesis,image enhancement and different target detections and so on.