Efficient Scene Text Detection Framework with Feature Pyramid and Expanded High-Level Feature Maps

Scene text detection is crucial for many applications but remains challenging due to large variations in aspect ratio, scale, and orientation.

This work proposes an efficient detection framework that combines an expanded high‑level feature map with a feature‑pyramid network (FPN) built on top of an SSD backbone. Text lines are decomposed into small, direction‑aware segments; a 8‑neighbor link predicts connections between segments, allowing reconstruction of arbitrarily long and rotated text.

Key components:

Interval sampling to enlarge high‑level feature maps, preserving resolution for small texts.

Fusion of deep and shallow features to construct a multi‑level pyramid (conv4_3_f, fc7_f, conv6_2_f, …) with 256‑dimensional channels.

Segment‑and‑link prediction on each pyramid level, modeling eight possible neighbor relations.

Geometric post‑processing that fits a line to linked segments and derives final bounding boxes.

Experiments on ICDAR2013 and ICDAR2015 show that expanding high‑level maps improves recall, while adding the pyramid further boosts precision. Compared with TextBoxes++, PixelLink and other state‑of‑the‑art methods, the proposed approach achieves a favorable trade‑off between speed (FPS) and accuracy.

The system is also deployed in a real‑world menu‑recognition scenario, where it outperforms SegLink by about 5 % on a 500‑image test set.

Future work will explore pixel‑level segmentation (inspired by PixelLink) and joint detection‑segmentation architectures.