Figure 5.

Overview of the main object-detection model. An input image containing the full reaction scheme is fed through a convolutional feature extractor (ResNeXt101), and the output feature maps are then passed to a Region Proposal Network, where regions of interest are selected and regressed. After being passed through fully connected layers, the learned class-specific information is used to classify and further regress bounding boxes to output image patches containing chemical diagrams, chemical labels, and arrow annotations.