Con-Text: Text Detection for Fine-Grained Object Classification.

Area of Science:

• Computer Vision
• Natural Language Processing
• Machine Learning

Background:

• Current fine-grained object classification methods primarily utilize visual cues.
• Integrating textual information from scene text remains an underexplored area in computer vision.

Purpose of the Study:

• To propose and validate a novel method for fine-grained object classification by combining visual and textual cues.
• To explore the efficacy of recognized scene text as a supplementary feature for object recognition tasks.

Main Methods:

• Textual cue extraction focusing on background regions and subsequent text recognition using OCR and character recognition algorithms.
• Encoding textual cues via bi- and trigrams with spatial constraints.
• Combining extracted visual and textual cues for classification and logo retrieval.

Main Results:

• Achieved a 15% improvement in end-to-end character recognition on the ICDAR03 dataset.
• Demonstrated that textual cues enhance fine-grained classification accuracy (70.7% vs. 60.3% with visual cues alone).
• Showcased the utility of textual cues for logo retrieval, improving accuracy from 54.8% to 57.4%.

Conclusions:

• Combining visual and textual cues offers a significant advantage over using visual information alone for fine-grained object classification.
• Recognized scene text provides valuable information for both classification and retrieval tasks.
• The proposed method for textual cue extraction and integration is effective and advances the state-of-the-art.