Cascaded Segmentation-Detection Networks for Word-Level Text Spotting

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Accurate text spotting in natural scenes is challenging due to variations in scale, illumination, and perspective.
• Existing methods often require post-processing steps like text line grouping, adding complexity and computational cost.

Purpose of the Study:

• To develop an efficient and accurate word-level text spotting algorithm.
• To eliminate the need for post-processing steps in text detection.

Main Methods:

• A two-stage convolutional neural network (CNN) cascade is proposed.
• The first CNN performs text region detection, providing coarse segmentation.
• The second CNN, inspired by YOLO, refines segments to predict oriented bounding boxes for individual words.

Main Results:

• The algorithm achieves accurate and reliable word-level text spotting in unconstrained environments ('in the wild').
• The system demonstrates good performance on the ICDAR 2013 and 2015 benchmarks.
• Execution time is efficient, at 450 ms for a 1000x560 image on a Titan X GPU.

Conclusions:

• The proposed cascaded CNN approach effectively addresses word-level text spotting.
• The system's ability to directly output word bounding boxes simplifies the text spotting pipeline.
• This method offers a promising solution for real-time text recognition applications.