SEMPANet: A Modified Path Aggregation Network with Squeeze-Excitation for Scene Text Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection frameworks are increasingly applied to text detection tasks.
• Scene text detection, especially for curved and small text, faces challenges with unbalanced data and feature extraction.
• Existing methods often rely on complex backbones like ResNet and FPN.

Purpose of the Study:

• To improve the performance of text detection in natural scenes, particularly for challenging cases like curved text and small targets.
• To develop a more efficient and lightweight text detection model.
• To enhance feature extraction capabilities in early stages of neural networks for text detection.

Main Methods:

• Utilized and improved the PSENet framework for text detection.
• Developed a novel SEMPANet (Single-stage, Efficient, Multi-path, Anchor-free) framework.
• Modified ResNet and FPN components for better early-stage feature extraction.
• Conducted experiments on ICDAR2015 and CTW1500 datasets.

Main Results:

• The improved network achieved a 1.01% higher F-measure on the ICDAR2015 dataset compared to PSENet-1s.
• The SEMPANet framework demonstrated superior performance over the original PSENet on the CTW1500 dataset.
• The proposed lightweight model achieved a training time of approximately 24 hours.

Conclusions:

• The enhanced PSENet framework, SEMPANet, is effective for scene text detection, especially for oriented and curved text.
• The modifications to ResNet and FPN improve feature extraction for text detection tasks.
• The proposed lightweight model offers a promising solution for efficient and accurate scene text detection.