Fig. 1 illustrates the overall architecture of the BFB-PTT method, which consists of a backbone network and a Bidirectional Fusion Branch (BFB), with sample space optimization through the PTT loss function. The BFB incorporates a short path fusion of low-level and high-level features, which gives it an advantage over traditional methods in terms of extracting and preserving low-level features. As shown in the diagram, the BFB connects various layers of ResNet50. The feature maps learned from