AAF-Net: Scene text detection based on attention aggregation features.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text detection faces challenges with small, multi-directional, and adjacent text instances due to neural network receptive field limitations.
• Existing methods often struggle with low detection rates and false positives for complex text arrangements.

Purpose of the Study:

• To propose a new feature pyramid network, the Cross-Scale Attention Aggregation Feature Pyramid Network (CSAA-FPN), for enhanced scene text detection.
• To address limitations in detecting small, arbitrarily oriented, and closely packed text instances.

Main Methods:

• Developed a novel Cross-Scale Attention Aggregation Feature Pyramid Network (CSAA-FPN).
• Incorporated an Attention Aggregation Feature Module (AAFM) to enhance features and handle multi-scale information.
• Utilized CBAM attention module for focused feature extraction and an Adaptive Fusion Module (AFM) for feature refinement.

Main Results:

• The proposed CSAA-FPN effectively enhances features, improving the detection of small and adjacent text instances.
• Experiments on CTW1500, Total-Text, ICDAR2015, and MSRA-TD500 datasets demonstrate the model's superior performance.

Conclusions:

• The CSAA-FPN model offers a significant advancement in scene text detection accuracy.
• The integration of attention mechanisms and adaptive fusion effectively tackles challenges posed by complex text scenes.