Research on Super-Resolution Processing of Improved RCAN on SMT Blurred Images

Surface mounted technology (SMT) of electronic motherboards is a key technology for assembling advanced circuit boards, and quality inspection is crucial for SMT assembled chips. Some blurred images in our data set cannot be directly used for training to obtain ideal detection results, so super-resolution processing needs to be performed first. We find that convolutional neural network depth and image feature correlation in training are critical for super-resolution. Therefore, we proposed the RCAN-CBAM network. The backbone network adopts the residual-in-residual structure of residual channel attention networks (RCAN) and builds a deep network through a combination of long and short residual connections to focus on high-frequency information learning. The original RCAN architecture only considers channel correlation and ignores spatial correlation. We introduce a spatial attention mechanism to form a convolutional block attention module (CBAM) to achieve comprehensive attention to channel and spatial features. Experiments show that our method shows better performance and visual effects than existing advanced methods on the SMT data set.