TL;DR
This paper introduces a novel three-stage approach for handwritten Chinese address recognition (HCAR) that significantly improves accuracy and efficiency. By combining coarse recognition for speed, recursive holistic word matching for robust recognition, and a modified edit distance for verification, the system outperforms existing methods on real-world datasets.
What Problem Does This Paper Address?
Handwritten Chinese Address Recognition (HCAR) is a challenging task due to the vast number of Chinese characters, variations in handwriting styles, distortions, and touching characters. Traditional methods often rely on segmenting individual characters, which is prone to errors. This paper aims to tackle these challenges by proposing a more robust and efficient system that can accurately recognize addresses from images.
Methods
The paper proposes a three-stage framework:
- Coarse Recognition: This stage aims to speed up the system by first extracting "key characters" (e.g., province, city, district) which are a small subset (22) of all Chinese characters. This reduces the complexity of the initial recognition task. It then filters potential "key character paths" based on address knowledge and matching distances.
- Fine Recognition: This stage employs a novel "recursive and enhanced holistic word recognition" method. Instead of segmenting individual characters, it recognizes entire "words" (place names) between key characters as a whole. The "recursive" aspect means it considers multiple candidates at each level and iteratively refines the recognition. This segmentation-free approach helps overcome segmentation errors.
- Verification: This stage further improves accuracy by using a "modified edit distance" to verify the results from the fine recognition stage, particularly for the last level of the address. This modified edit distance accounts for both the holistic word recognition results and traditional character recognition candidates.
Datasets
The paper evaluates its approach on three real-world datasets: low quality, medium quality, and good quality. These datasets consist of 854 images in total, collected from real users. The paper does not explicitly state if these are newly introduced benchmarks or existing ones.
Key Insights and Ideas
- Key Character Focus: Reducing the initial recognition problem to a small set of "key characters" is a clever way to improve efficiency and accuracy in the first stage.
- Segmentation-Free Holistic Recognition: Recognizing entire place names as holistic units, rather than relying on character-by-character segmentation, is crucial for handling handwriting variations and touching characters.
- Recursive Candidate Evaluation: Considering multiple candidates at each recognition level and recursively evaluating them significantly boosts accuracy, especially when place names are visually similar.
- Verification as a Final Polish: The modified edit distance provides a powerful mechanism to correct errors in the final recognition stage by leveraging both holistic and character-level information.
Key Results
The proposed three-stage system (Our Appr.+Veri.) achieved the highest String Recognition Rate (SRR) across all three datasets:
- Low Quality: 71.74%
- Medium Quality: 80.70%
- Good Quality: 92.54%
Compared to the previous holistic word matching approach (HWA), the proposed system showed significant improvements: 15.94% for low quality, 14.38% for medium quality, and 6.76% for good quality data. The coarse recognition stage was shown to reduce processing time by at least half without significantly impacting accuracy. The trimming-down strategy further halved the average recognition time.
Limitations and Future Work
- Limitations: The paper acknowledges that the plain recognition approach (PRA) performs poorly due to segmentation errors and handwriting variations. While their system significantly improves upon this, the accuracy on low-quality data still leaves room for improvement. The paper also mentions that the "plain recognition is highly unreliable" in certain scenarios, necessitating the holistic approach.
- Future Work: Not explicitly detailed in the provided text, but the success of the three-stage approach suggests further refinement of each stage and potentially exploring more advanced feature extraction or deep learning methods for holistic recognition. The paper's focus on efficiency also hints at continued efforts to optimize processing speed.
Summary updated