Font Size: a A A

Study On High SNR Narrow-band Spectral Data Acquisition And Unsupervised Classification

Posted on:2015-05-31Degree:DoctorType:Dissertation
Country:ChinaCandidate:J YueFull Text:PDF
GTID:1108330482969721Subject:Physical Electronics
Abstract/Summary:PDF Full Text Request
Compared with traditional imaging detection, hyperspectral data contains both spatial and spectrum information of scene, and hyperspectral data can describe the scene more precisely and richly. It is much more efficient than traditional imaging for target detection. In order to describe the object more precisely, bandwidth of modern hyperspectral imager is about several nanometers. When the source tends to be weak the SNR of acquired data would be poor due to such narrow band, so how to acquire high-throughput and high SNR hyperspectral data with such narrow bandwidth is a valuable problem. When hyperspectral data has been acquired, classification is needed to identify targets. High performance unsupervised classification is a fine choice for target detection due to none of effective prior knowledge of targets. In conclusion, this paper focuses on high SNR hyperspectral data acquisition and high performance unsupervised classification. The main contributions and innovation points of this paper are listed as follows.(1) High SNR spectrum acquisition based on double light path. Hadamard transform spectrometer (HTS) needs long measurement-period-time and high-precision encoding components. A novel high-SNR spectrometer based on double path is proposed to overcome these shortages of HTS. An optical convolution structure is implemented to measure fast-changing spectrum. This proposed method has two main advantages: simplified the optical structure by eliminating the coding component, improved measurement speed due to only one-time measurement needed.(2) Study on denoising performance analysis of Hadamard transform spectrometry in theory and experiment. The pre-conditionals of traditional Hadamard transform spectrometry denoising analysis theory are not always met, and the existing conclusions are conflictive. Denoising of HTS is re-analyzed via new noise model based on classification of spectrum. The results support the existed theory that HTS outperforms traditional spectrometer. The conclusions have extended the established HTS theory.(3) Sparse signal reconstruction in Hadamard matrix coding measurement. SNR of HTS not only depends on noise environment but also depends on the structure of signal, such as the signal is sparse. The correlation of noises in restored signals and measured coded signals has been analyzed. The results show that noise in reconstructed signals related with the average noise value of measured coded signals. Sparse signal reconstruction method is employed to restore the sparse spectrum. The results show that HTS outperforms slit-based spectrometry when the signal is sparse, and the sparser the signal the higher SNR of reconstructed signal will be.(4) Unsupervised clustering based on spatial coherence property and salient target extraction. Most of unsupervised clustering methods need the number of targets which is not available in reality. Besides, these methods prefer spectrum information to spatial information for clustering. Two novel unsupervised clustering methods utilizing spatial coherence property are proposed:unsupervised classification based on min-related-window and unsupervised classification based on pixel-reducing, those two methods aim to solve noise, edge fuzziness and target incomplete. According to that targets can been detected in several specific bands, a salient target extraction method based on edge and transition region is proposed to solve time consuming.(5) Narrow-band imagine system and camouflage net detection experiments. There are two narrow-band imaging spectrometers proposed in this paper, component-based transmissive imaging spectrometer and dual-dispersive imaging spectrometer. Based on these two designs, camouflage net detection experiments have been implemented. The results show that camouflage net and grass can be distinguished in several specific bands, but not all the targets. Furthermore, the proposed clustering method is employed to detect targets, and the result show that all targets can be detected clearly.
Keywords/Search Tags:Target Detection, Narrow-Band Image, Signal to Noise Ratio, Hadamard Transform, Sparse Signal, Unsupervised Clustering
PDF Full Text Request
Related items