FIG 3.

Schematic representation of the GFS-ExtremeNet algorithm in microscopic-image detection. Overall, this geometry-aware approach is divided into two parts. (A) The first part is using the hourglass network to extract the extreme and the center points of the target and generating heatmaps for those points. In addition, the relationship of the extreme points is measured and passed to the second stage in forming the adjacency spectrum. (B) The second part is the feature combination, mainly center grouping. (C) After the geometric relationship is combined, the adjacency spectrum is then used for verification to form the detection results.