Research On Key Problems Of Anomaly Target Detection In Active Hyperspectral Images

Hyperspectral anomaly detection technology refers to the identification of abnormal pixels under the circumstance that background and target prior information are missing,where the abnormal pixels are different from the background spectral pixels.It is widely used in vegetation research,water quality detection and military reconnaissance and other fields.Thus the hyperspectral anomaly detection has significant meanings in army and livelihood.In order to achieve accurate identification of anomalous targets,this task requires not only high-performance detection technology,but also reliable acquisition equipment that provides high-quality data.However,in the environments with weak sunlight,i.e.,the night or extreme environmental conditions,the traditional hyperspectral imaging systems fail to identify the abnormal pixels due to the poor lighting conditions.Poor lighting conditions makes it impossible to locate the target in the scene,and it is difficult to perform anomaly detection tasks.Due to the above reasons,there are few researches anomaly detection technology under the poor lighting conditions.To tackle the problems,this dissertation proposes to use supercontinuum lasers,short-wave infrared spectroscopy imager,two-dimensional turntables and laser shaping systems to build an active hyperspectral imaging system.Besides,this dissertation completes hyperspectral data collection,which provides the reliable data for subsequent hyperspectral anomaly detection under the poor lighting conditions.A series of denoising and anomaly detection algorithms are proposed for the problems of active hyperspectral images.The main research content of this dissertation consists of the following three aspects:1.For the signal enhancement requirements of active hyperspectral images with low signal-to-noise ratio,this dissertation proposes an active hyperspectral image restoration technology based on subspace mapping.This dissertation uses the convolution network to learn a set of orthogonal bases on the feature map independently,and projects the low-dimensional feature maps to the subspace to suppress the noise signals.Then,an up-sampling operation is performed on the feature map after projection to restore a complete noise-free image.In addition,compared with the traditional distance measurement method,this dissertation uses convolution operations to optimize the distribution of denoising images to approximate the distribution of high-quality images.Compared with the algorithms with the best denoising results,the NRHQ value of this algorithm on three datasets has dropped by 22%,41%,and 32% respectively.The algorithm in this dissertation can achieve the best compromise between preserving image details and removing background noise.2.To solve the problem of the high false alarm rate of the detection result caused by the mixture of background and abnormal pixels in the active hyperspectral images,this dissertation proposes a detection method,which is a deep convolution neural network with weight adjustment strategy.The traditional anomaly detection methods ignore the non-linearity and complexity of the hyperspectral image,which making less use of spatial information.This dissertation leverages three-dimensional convolution operation instead of the two-dimensional convolution to model the high dimensions of hyperspectral data.Besides,the method in this study generates the weights for the features automatically by measuring the absolute distance of the pixels and the angle of the spectrum.The values of the weights indicate the similarities between the pixels,and the small values means that the more similarities of the pixels.Compared with other traditional detection algorithms,this algorithm strengthens the separability between background and abnormal pixels,and improves the detection performance under low false alarm rate.3.To solve the problem that low spatial resolution leads to low accuracy of active hyperspectral abnormal target detection,this dissertation proposes a super-resolution reconstruction anomaly detection algorithm based on attention mechanism.The algorithm consists of two main models: a super-resolution reconstruction model and active hyperspectral anomaly detection model.The super-resolution reconstruction model selectively models global spatial spectrum information,focusing on enhancing the spatial characteristics of the area where anomalous pixels are located,while suppressing background features far away from anomalous targets.Besides,the active hyperspectral anomaly detection model then detect the abnormal targets in the reconstructed hyperspectral images with the high-frequency.The algorithm combines the super-resolution reconstruction task together with the anomaly detection tasks to assure that the reconstructed images satisfy the requirements of the detection task.These two models are optimized in turn to promote each other to improve the performance on hyperspectral abnormal target detection under the low-resolution scenerio.On the measured active hyperspectral data set,compared with the other five representative algorithms,the average detection accuracy of this algorithm is improved by 13%.