We propose a method for a SMD segmentation by using a convolutional encoder-decoder. Our method distinguishs an information and a location of a SMD on the PCB from an optical image. We use the encoder-decoder architecture. The encoding network is based on VGG net. The decoding network is composed of a convolutional transpose and a convolution layer. The proposed method achieves 98.8％, 96.8％ precisions about the register and the capacitor.