In tradition, surface mounted devices (SMD) are classified and registered manually in surface mount technology (SMT). Manual classification is inefficient and relies on the experience of the operators. We propose a novel SMD classification method based on a mixture neural network. The network contains two modules, one is a convolutional neural network with a gray image as input, and the other is a fully connected network with a Hand-Crafted feature as input. The lighting environment of chip has a great impact on chip imaging, and the classification method should maintain sufficient robustness against light changes in SMT. However, there is a lack of a robust feature extraction method for SMT classification against illumination changes. A novel feature extraction method based the characteristics of the binary chip image is first proposed, which is composed of the mathematical statistical pin parameters of SMD. As the light changes, the binary image is relatively stable, making the proposed feature is invariant to light changes. An SMD dataset in SMT contains 61 kinds of chips is built for training and testing. Experiments on this dataset demonstrate that our method is more effective than the state-of-the-art methods in terms of performance measures.

传统上，表面贴装器件 (SMD) 在表面贴装技术 (SMT) 中手动分类和注册。人工分类效率低下，依赖于操作人员的经验。我们提出了一种基于混合神经网络的新型 SMD 分类方法。该网络包含两个模块，一个是一个以灰度图像作为输入的卷积神经网络，另一个是一个以 Hand-Crafted 特征作为输入的全连接网络。芯片的光照环境对芯片成像影响很大，分类方法应对SMT中的光照变化保持足够的鲁棒性。然而，缺乏针对光照变化的 SMT 分类的鲁棒特征提取方法。首次提出了一种基于二值芯片图像特征的特征提取新方法，由SMD的数理统计管脚参数组成。随着光线的变化，二值图像相对稳定，使得提出的特征对光线变化具有不变性。SMT 中的 SMD 数据集包含 61 种芯片，用于训练和测试。在这个数据集上的实验表明，我们的方法在性能测量方面比最先进的方法更有效。