GS-U-Net-An Catalyst Image Segmentation Model

The GS-U-Net presented in this study is a lightweight deep-learning model designed to improve the characterization of nanoparticles using transmission electron microscopy (TEM). The GS-U-Net is an innovative adaptation of the U-Net architecture featuring a VGG16-based encoder with a squeeze excitation (SE) attention mechanism and a GhostBottleneck module for enhanced feature extraction. The decoder part also integrates the SE mechanism and utilizes Ghost Convolution (GhostConv) instead of the standard convolution, which improves detection accuracy and reduces computational requirements. The performance of the model is evaluated by its Mean Intersection Union (MIOU) score, which improves from 89.71% to 90.07%. In addition, the parameter size of the model was drastically reduced from 24.892MB to 15.563MB, which significantly reduces the computational load and increases the detection speed. The improvement in efficiency and accuracy of the model helps to advance nanoparticle characterization research.