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

Table 3

Comparison results on Total-Text. “,” “,” and “” represent “Precision,” “Recall,” and “-score”, respectively. The best and second-best terms are highlighted in italic and bold, respectively.

Method	(%)	(%)	(%)
TextSnake [15]	82.7	74.5	78.4	—
ATRR [25]	80.9	76.2	78.5	—
MTS [14]	82.5	75.6	78.6	—
TextField [26]	81.2	79.9	80.6	—
LOMO [27]	87.6	79.3	83.3	—
CRAFT [28]	87.6	79.9	83.6	—
CSE [29]	81.4	79.1	80.2	—
PSE-1s [16]	84	78	80.9	3.9
DB-ResNet-18 [1]	88.3	77.9	82.8	50
DB-ResNet-50 [1]	87.1	82.5	84.7	32
The proposed (Pr0.1)	86.5	80.4	83.3	52
The proposed (Pr0.2)	86.6	80	83.2	54
The proposed (Pr0.5)	86.9	79.1	82.8	57