Adapted Fast R-CNN diagram, where convolutional layer is abbreviated as ‘Conv layer’, fully connected layer is abbreviated as ‘FC layer’, and region of interest is abbreviated as ‘RoI’. In the original Fast R-CNN, Selective Search (Uijlings et al. 2013) is a general way to generate region proposals by exhaustively sampling regions in any image based on hierarchical image segmentation. Instead, we pre-computed our own region proposals as source detection software provides us with the exact locations of significant blobs of radio emission (see Sect. 4.1). This means we can also disable the part of the Fast R-CNN that is designed to refine the location and dimension of proposed regions.