Extreme Learning Machine Based Defect Detection for Solder Joints

The quality of solder joints is essential for electronic products, and the detection of defects in solder joints is critical to the quality control of electronic products. A vision inspection is developed to detect defects of solder joints in automatic line. Extreme learning machine is applied to identify defective solder joints from qualified ones. Five low level features and three advanced features are employed as input features. The low-level features include roundness, roughness, entropy, contrast and histogram of oriented gradient. The advanced features include grey-level co-occurrence matrix, local binary pattern, and segmentation-based fractal texture analysis. To solve unbalanced samples problem, Gaussian mixture model based dense estimation scheme is proposed to adjust the classification super plane for extreme learning machine. The experimental results demonstrate that the proposed defect detection method is more efficient than neural network, support vector machines, common extreme learning machine and convolutional neural network-based methods, and it has real-time performance to meet the equirement of the actual production line.