在多目标光纤光谱望远镜系统数据处理中，二维光纤光谱抽取是非常重要的组成部分。天文学家通过抽取的光谱来获取天体的各种物理信息，光谱抽取精度对后续的光谱识别和分析有直接影响。根据成像原理，二维光纤光谱的抽取是一个反卷积的过程。在该过程中，点扩散函数（Point Spread Function, PSF）的精确估计和快速反卷积方法是实现二维光纤光谱反卷积光谱抽取的关键所在。点扩散函数估计得越精确，则光谱抽取的效果就会越好；反卷积算法越快，则能进行实际应用。因此，本论文的研究内容主要包括两部分：点扩散函数模型和快速反卷积抽取方法。其中，点扩散函数模型研究分为三个方面：大气湍流和望远镜系统的点扩散函数、光纤和光谱仪的点扩散函数以及空间变化的点扩散函数。大气湍流和望远镜系统的点扩散函数模型，是目标天体的光从大气层进入望远镜系统（尚未进入光纤）所形成的点扩散函数。本论文通过傅里叶变换所生成的模拟随机相位屏对大气湍流的数值模拟来计算短曝光点扩散函数的畸变，并采用高斯过程来表示短曝光PSF移动的特征，提出一种新的短曝光PSF基于高斯过程移动叠加形成长曝光PSF的估计方法。光纤和光谱仪的点扩散函数模型，可以通过定标灯光谱上的斑点来估计。为了表示各种具有不同径向扭曲和椭圆率的斑点，本论文提出了一种改进的二维高斯点扩展函数模型来拟合定标灯光谱的斑点成像，并分别采用最速下降法、拟牛顿法和改进的和声粒子群优化算法来优化所提高斯模型的最佳参数。天体目标在二维光纤光谱上成像时，点扩散函数的位置，以及形状和幅度，是空间变化的，它随着图像空间位置不同而不同，这是由光谱仪存在的像差和仪器中各种元件的不同步震动引起的。针对该问题，发展了空间变化的点扩散函数估计方法。首先，提出了一种基于高斯低通滤波的光纤轨迹中心追踪方法，来确定了光纤色散方向上某些指定位置的光纤轨迹中心，即点扩散函数中心位置。其次，采用拉格朗日多项式、牛顿多项式和三次样条多项式，这三种多项式插值算法求出其它空间变化的点扩散函数中心位置，并将二维点扩散函数转换成一维向量形式，分别采用上述三种多项式进行插值，估计出空间变化的点扩散函数的形状和幅度。最后讨论了空间变化的点扩散函数所形成的降晰矩阵的两种构造方法：图像分块策略和降晰矩阵分解法，分析了空间变化的降晰矩阵的反卷积运算复杂度。从二维光纤光谱图像中抽取出有重要科研价值的天体目标一维光谱，即二维光纤光谱的光谱抽取，可利用反卷积方法进行抽取。针对降晰矩阵的稀疏、非对称和规模巨大的特点，采用最小二乘QR分解法，即LSQR方法进行求解，解决了目前反卷积抽取方法由于超大降晰矩阵难以计算而无法投入实际应用的问题。为了进一步阻止LSQR算法早熟收敛导致求解精度受影响，并抑制反卷积算法中存在的“振铃”现象，提高反卷积光谱抽取的抗噪声能力，提出了自适应正则化LSQR算法。另外，给出了一种自适应的正则化参数选择策略，可以很方便的确定正则化参数。研究了将反卷积方法与基于GPU的并行计算技术有效整合，进一步加快提出方法的运算速度。通过在真实的LAMOST二维光纤光谱和模拟二维光纤光谱上进行的一系列实验，验证了提出的反卷积抽取方法的有效性，并给出与其它方法量化比较的结果。

The extraction of two dimensional fiber spectroscopy is an important step of data processing in multi-object fiber spectroscopy telescope system. Physical information of astronomical objects is achieved by astronomer. And the precision of spectrum extraction has direct effect on subsequent spectrum recognition and analysis. According to imaging theorem, the extraction of two dimensional fiber spectroscopy is a process of deconvolution. During the process, accurate point spread function estimation (PSF) is the vital prerequisite of realization of spectrum extraction with fast deconvolution method. The more accurately the PSF estimates, the better result the spectrum extraction obtains. The faster the deconvolution method runs, the more practical the algorithm acts. Therefore, this thesis mainly includes two parts: the imaging PSF of two dimensional fiber spectroscopy and fast deconvolution extraction method. Among in the study of the PSF, the discussion is divided into three issues: the PSF of atmospheric turbulence and telescope system, the PSF of fiber and spectrograph and spatially-variant PSF. The estimation of the PSF of atmospheric turbulence and telescope system is to study the PSF after the light of astronomical objects entering the atmosphere and telescope system (still not geting into fiber). Study through the turbulence screen by FFT method to simulate the atmospheric turbulence for the calculation of aberration of short exposure PSF. With representation of the movement characters of short exposure PSF, a novel long exposure PSF estimation based on short exposure PSF Gaussian processing movement and superposition is proposed.The PSF of fiber and spectrograph can be estimated by the speckle of the arc-lamp image. An improved two dimensional Gaussian PSF model is proposed to fit the arc-lamp speckle. The proposed model can represent different radial distortion and ellipsicity. While Gaussian function is non-linear function, non-linear optimization algorithms, including steepest descending method, quasi-newton method and proposed improved harmony PSO algorithm are used to fit the best model parameters.The spatially-variant PSFs represent the centroid position, shape and scope being slightly changed with the variation of the space position for being affected by the aberration, the asynchronous vibration of the components in the instrucment, while astronomical objects imaging on the two dimensional fiber spectrum image. A fiber centroid searching method based on Gaussian lowpass filtering is proposed to some specified fiber centroid position, i.e. PSF centroid position. Lagrange polynomial, Newton polynomial and cubic spline polynomials are used to interpolate for the calculation of other spatially-variant PSFs centroid position. Moreover, two dimensional PSF function is changed into one dimensional formation and interpolated by this three interpolation polynomials, for estimating the shape and scope of the spatially-variant PSFs. The corresponding construction methods of blurring matrix by spatially-variant PSFs are given out, including image partitioning tactic and blurring matrix decomposition method. The time complexity of deconvolution method for spatially-variant blurring matrix is analysized. The deconvolution extraction method of two dimensional fiber spectrograph is studied, i.e. to extract one dimensional spectrum with scientific values of astronomical objects from two dimensional fiber spectroscopy. Aimed at the sparse, non-symmetrical and large-scale features of the convolution matrix, the least squares QR- factorization (LSQR) algorithm is used to calculate, which solved the non-practical problem of deconvolution extraction method due to huge scale convolution matrix. Moreover, for preventing early convergence and restraining the ringing phenomenon of deconvolution algorithm and the improvement of anti-noise, adaptive regularization LSQR algorithm is proposed. In addition, a kind of adaptive regularization parameter selection strategy is given out, which can conveniently choose the parameters. Moreover, combining the GPU paralleled calculation technology with proposed method is studied for further acceleration. The proposed is validated, through a serial of experiments on actual LAMOST two dimensional fiber spectrograph and simulation fiber spectrograph, and the results of quantization comparison with other extraction method.