Fig. 9.

The architecture of region localization stage (RLS). In this figure, ‘$\text{Conv}f\times f,c,/s$’ denoted convolutional layer with size of filters f, and number of channels c and strides s (default strides was 1). Noting that each Conv layer was followed by a BN and an activation layer of ReLU. ‘$\text{Pool}f\times f,/s$’ meant max pooling layer whose size of filters f and strides s. ‘$\text{Up-sample},/r$’ indicated nearest neighbor up-sampling with up-sampling rate r. ‘Anchor’ was anchor box which was utilized to predict the PCoA region. The first part was an input receiving a three-channel RGB image. The following feature extraction block was FPN with backbone of ResNet50. At last, the anchor boxes output the PCoA region of original input image