Sparse Reconstruction Method Of Micro Focus X-ray Image Based On Learning

The high precision industrial Computed Tomography(CT) technology has attracted great interest in defect detection of manufacturing process, such as integrated circuit(IC), power insulators, aerospace device and other high precision device. With the requirements of high resolution, X-ray imaging technique with micro spot size has recently received much attention in high precision industrial CT, namely micro-focus X-ray imaging technology. However, compared with general images, micro-focus X-ray images contain a complex mixture of noise, and have low contrast and signal to noise ratio(SNR). While general reconstructing algorithms are normally focused on the single noise model which is not suitable to the micro-focus X-ray images.Based on dictionary learning, sparse reconstructing theory can achieve the effect of image reconstructing and improving SNR by restoring sparsity of the images. Compared with other traditional reconstructing algorithms, sparse reconstruction is more suitable for micro-focus X-ray images which are sparse naturally. On the basis of specific characteristics appearing in micro-focus X-ray images, this paper concentrates on sparse reconstructing algorithms based on dictionary learning for Poisson-Gaussian noise model. The main research items are as follow:(1) Based on image sparse reconstructing object function for mixed noise, the sparse reconstructing object function in which the data fidelity term and sparsity constraint term are separating, and a sparse reconstructing algorithm combining K-singular value decomposition(K-SVD) and total variation(TV) regularization is proposed. By analyzing the process of the micro-focus X-ray imaging, this paper establishes the sparse reconstructing object function in line with the mixture model of Poisson-Gaussian noise, and proposes a sparse reconstructing algorithm with K-SVD and TV regularization. Such algorithm adaptively adjusts fidelity coefficient and use K-SVD and TV regularization interchangeably to reconstruct and optimize the reconstructing results. Finally, the correctness of Poisson-Gaussian mixture model and the effectiveness of the reconstructing result of the algorithm can be verified by some experiments in this paper.(2) Based on image sparse reconstructing object function for mixed noise, the sparse reconstructing object function in which the data fidelity term and sparsity constraint term are integrating, and a sparse reconstructing algorithm is proposed for the strong Poisson-Gaussian mixture noise. By representing the images used exponent form, block clustering algorithm is firstly used to reduce the amount of calculation of dictionary learning, and then greedy algorithm is employed to decompose image in which the data fidelity term and sparsity constraint term are integrating, while the dictionary update is turned into optimization problem which can be solved easily. Experiment results show that this algorithm has a more natural detail and is more adaptability for strong Poisson-Gaussian mixture noise compared to the general image reconstructing methods.