Research On The Key Image Processing Problems Of X-ray Detection For Ultra Large-scale Integrated Circuit Packaging

Ultra large scale integration circuit(ULSIC) is one of the important pillars for the national economy and the construction of national defense, which is also an essential index of the scientific and technological development for a country or an area. With the development of multi-function, high density and multi-layer packaging in the integrated circuits, some existing automated optical inspection methods are merely applicable for surface defects, whereas they are infeasible for the internal defects in multi-layer packaging. The imaging process of X-ray detector can penetrate packaged components and observe the internal defects directly, which provides an effective means to tackle internal defect detection for ULSIC. However, due to the fact that most of packaging components are metal material with features of high density and small size, X-ray images exhibit characteristics of low signal-to-noise ratio, low contrast and contain very small fine features, resulting in extreme challenges for internal defect detection. To this end, the research on image restoration methods for internal defects detection in X-ray images appears to be of great significance for both theoretical and applicable value.This thesis researches on regularized image restoration technologies of X-ray detection for ultra large scale integration circuit packaging, which is mainly based on sparsity based 2 1l-l model and total variation(TV) regularization model, involving new fast algorithms and improvements of the models from the viewpoints of physical motion and derivative space. The main research items are as follows:1) An accelerated gradient descent with momentum based method for image deblurring is proposed. This method is designed to eliminate the drawbacks of conventional gradient descent direction in 2 1l-l model based image restoration problems, such as slow convergence and high sensitivity to noise. This method simulates the concepts of momentum in physics and iterates along the direction in gradient descent with momentum. By adding a momentum term to negative-gradient direction, this search strategy allows to take larger steps in favorable directions and to prevent the sudden change of descent direction caused by noises, resulting in accelerated convergence and overstepped local optima. The necessary and sufficient conditions for the existence of a stable solution has been theoretically presented and proved. Two sets of experiments on standard gray images and X-ray images demonstrate the effectiveness of proposed algorithm in restored quality and rapidity.2) In accordance with the proximal theory, a convex quadratic approximation model for a class of non-differentiable regularization term based restoration model is derived in the framework of momentum gradient descent. Combined with non-differentiable total variation(TV) regularization, a momentum based gradient projection method is proposed for TV image restoration problems, and the convergence is theoretically proved. This method can avoid the computational difficulties caused by non-differentiability TV regularization. Finally, the experimental results demonstrate the superiority of the proposed method in computational speed and the effectiveness in terms of removing the noise and preserving the details.3) From the viewpoint of physics motion, the iteration process of sparse reconstruction algorithms is initially investigated from the perspective of particles’ moving process. In this paper, a physical total energy based objective function model is proposed on the basis of 2 1l-l model. Then, a targeted gradient projection technique is adopted to solve such a reconstruction model, and its convergence is discussed as well. In the process, the particles’ motion model is created in viscous medium and Newtonian fluid with the particle’s gravity potential energy function denoted on the basis of 2 1l-l model. Furthermore, in discrete calculations, the particle’s displacement is defined by the corresponding iterative result. Compared to five state-off-the-art methods, the experimental results demonstrate that the proposed method outperforms its competitors distinctly in time efficiency on the basis of guaranteeing the reconstruction quality, and is more suitable for applications to X-ray defect detection4) Based on the connections between image space and derivative space, the novel anisotropic TV and isotropic TV models based on alternated direction gradients are derived in derivative space. Then, the related anisotropic and isotropic algorithms are proposed based on split bregman frame. These models can improve the effectiveness of traditional fidelity term because the derivative space is useful in improving the success rate of image restoration. These methods decouple the horizontal gradient and vertical gradient in derivative space, which is of advantage to improve the ability of denoising and detail preserving. Finally, experimental results show that the proposed method can obtain satisfactory restoration results, and is superior in denoising and detail-preserving to the TV regularization method in image space.