Research On Key Techniques For Semiconductor Chip Surface Defect Online Detection Based On Machine Vision

Semiconductor chips are widely used in many fields and all kinds of electronic products, and have been the lifeline of economic development and national information security, profoundly affecting the modern people's life. During the manufacturing and encapsulation process, defects appear on surface of the chips inevitably, which will affect the chips' operation efficiency and life span. Traditional artificial visual methods are unable to meet the requirements of high speed and high accuracy in inspection. Inspection on surface of chips using machine vision has the advantages of non-contact, non-invasion, high accuracy, high speed, and high stability. Although significant progresses have been made on defect detection of printed characters and appearance,size,location of leads based on machine vision, researches on appearance defect detection and classification on chip surface are still at the early stage. Furthermore, higher demand for image processing algorithms' feasibility, accuracy, real-time, efficiency in semiconductor chip surface defect online detection and classification are put forward. Therefore, this thesis does the researches on key techniques for?semiconductor chip surface defect online detection based on machine vision, establishes an integrated online detection theoretical system about semiconductor chip image acquisition, image pre-processing, image segmentation, surface defect feature extraction and classification. This thesis will be helpful to the development of key techniques for semiconductor chip surface defect online detection, with important theoretical research sense and extensive application value. The main research are as follows:(1) Design of chip surface defect online detection system. Firstly, the overall structure of detectionsystem is designed, and the basic working procedure is described. Secondly, based on the characteristics of leads, central pad and plastic encapsulation packages,with the requirements of surface defect detection, a novel bright-dark-field light source design for semiconductor chip image acquisition is proposed, to complete the image acquisition task in high dynamic range in one time exposure. Thirdly, key detection components and their working principle are studied, camera and lens are selected based on detection requirements and parameters calculation. The mathematical model for the depth field (DOF) of image acquisition system is established. The system operation parameters in application can be adjusted by the DOF model. Fourthly,software modules and implementation process are designed, including offline training and online detection and classification. Lastly, the lead images and plastic encapsulation package images (hereafter referred to as package images) in high quality are acquired efficiently using the proposed semiconductor chip surface defect online detection system.(2) Research on method of chip surface defect image pre-processing. Firstly,fast image separation and correction algorithms for both lead images and package images are proposed, aiming at separating all single chip images from the images captured by the system,rotating and optimizing all single chip images automatically. Secondly, in order to enhance image resolution,highlight the defect characteristics and preserve edge information, based on the implementation of parallel bicubic interpolation algorithm, an enhanced fast edge directional cubic convolution interpolation algorithm is proposed. Otsu function is used as reference to determine the edge type and direction in local window. The missing pixel on strong edge is calculated using cubic convolution along the edge direction, while the missing pixel on weak edge or in texture region is calculated by combining two orthogonal directional cubic convolution. The experimental results prove that the proposed image interpolation algorithm outperforms the competitors in terms of preserving edge and texture information.(3) Research on method of chip surface defect image threshold segmentation. Firstly, based on the characteristics of chip surface defect images, multilevel image thresholding algorithm based on fuzzy entropy and modified gravitational search algorithm is proposed, in which trapezoidal membership function is adopted to transform the multilevel image thresholding problem into optimizing fuzzy parameters of multilevel fuzzy entropy. The original gravitational search algorithm is modified to accelerate the multilevel fuzzy entropy maximization process. Experimental results prove the effectiveness and superiority of the proposed algorithm. However, this algorithm can hardly be applied to the situation with high real-time demand due to the large time consumption of parameters' calculation. Secondly, in order to enhance the global search ability and the convergence ability of optimization algorithms for multilevel image thresholding,a novel hybrid particle swarm optimization and gravitational search algorithm with generalized opposition-based learning (GPSOGSA) for multilevel image thresholding is proposed. In GPSOGSA, the modified gravitational search algorithm is combined with particle swarm optimization, together with strategies of generalized opposition-based learning and normal mutation on the global best agent. The multilevel Otsu function is used as the criterion of multilevel image thresholding. Computational complexity analysis is performed and the experimental results show that the proposed GPSOGSA outperforms the competitors in terms of population diversity, global and local search ability and convergence ability. The optimal parameters for multilevel image thresholding of lead images using GPSOGSA are carried out. Lastly, multilevel thresholding algorithm of 2D maximum entropy based on gray level-gradient co-occurrence matrix is proposed to segment package images. The gray level and gradient information are both considered and GPSOGSA is used to optimize the 2D entropy maximization process. The segmentation results of the proposed algorithm are better that those of Otsu with single threshold, and comparable to the exhaustive search which is much slower than the proposed algorithm.(4) Research on method of chip surface defect feature extraction and classification. Firstly, five types of package defect and five types of lead defect are concluded and analyzed based on actual chip appearance inspection requirements and segmented images. Secondly, the package defect region extraction method based on mathematical morphology is proposed. Grayscale images of package defect region are obtained. the lead defect region extraction method based on texture direction is proposed, in which defect regions are judged and located based on the texture direction information. Grayscale images of lead defect region are obtained.Thirdly, based on the grayscale images of defect regions, 24 types of features including geometry, gray and texture features are extracted. Fourthly,in order to eliminate irrelevant or redundant features and enhance the classification efficiency,support vector machine recursive feature elimination (SVM-RFE) algorithm is used to perform feature selection on the defect feature samples and the optimal feature subsets are obtained. Defect feature samples that only contain the optimal feature subset are then obtained. Lastly, the radial basis function support vector machine (RBF-SVM) based on GPSOGSA is proposed for defect classification, in which the multi-class classifier is designed based on SVM one versus one binary classifier, and the search process for penalty factor and the parameter of Gaussian kernel is optimized by GPSOGSA. Experimental results shows that GPSOGSA can obtain comparatively better penalty factor,the parameter of Gaussian kernel and the classification accuracy of 10-fold cross validation.(5) Experimental validation and analysis on chip surface defect online detection. Firstly, the technical indexes of online detection are put forward, and the experimental environment of semiconductor chip surface defect online detection is established. Secondly, after achieving complete process from single chip images separation and correction, image interpolation, image threshold segmentation, defect region extraction, defect feature extraction, to final defect classification, the real-time and feasibility of the proposed defect online detection algorithms are validated and analyzed. The experimental results meet the demand of all the online defect detection technical indexes.