An Agile and Efficient Neural Network Based on Knowledge Distillation for Scene Text Detection

Table 2

The experiment results with different losses. Pr: the model is obtained by pruning (the pruning ratio is 0.1); Ft: the model is fine-tuned; Fl: feature loss is adopted; Pl: probability loss is adopted. “,” “,” and “,” represent “Precision,” “Recall,” and “-score,” respectively.

Method	Pruning ratio	(%)	(%)	(%)
DBNet	—	88.3	82.5	82.8
Pr	0.1	83.4	80	81.1
Pr+Ft	0.1	85.5	80.1	82.7
Pr+Ft+Fl	0.1	86.4	79.8	83
Pr+Ft+Fl+Pl	0.1	86.5	80.4	83.3