Back to papers
    OCRtext recognitionCRNNCTCLSTM

    An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition

    Baoguang ShiXiang BaiCong Yao
    📅 2017🏛 IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 11, pp. 2298-2304, DOI: 10.1109/TPAMI.2016.2646691
    Problem

    This paper addresses the challenge of text recognition in images (especially natural scenes) with varying fonts, sizes, and orientations by introducing a fully trainable end-to-end network.

    Method

    The CRNN architecture combines a CNN for visual feature extraction, a bidirectional RNN (LSTM) for sequence modeling, and a CTC layer for mapping feature sequences to character outputs without requiring character-level segmentation.

    Finding

    The proposed method achieved competitive and superior accuracy on standard benchmarks like IIIT-5K and ICDAR, while being end-to-end trainable and requiring no manual segmentation.

    Limitations

    Performance degrades on very long texts, texts with large character spacing, and languages with special characters (e.g., Arabic), and the model requires large training datasets.

    Practical application

    CRNN is used in automated registration systems, license plate reading, and document scanning. In the finger robot project, this model is ideal for recognizing text on the screen (button labels, window titles, text content) so the robot can read on-screen instructions and act accordingly (e.g., click 'Save' or type in a search field).

    📇 Summary flashcard — 13 analytical fields for this paper

    خلاصه

    CRNN presents an end-to-end trainable network for sequence recognition that combines CNN, RNN, and CTC to recognize variable-length text without segmentation.

    نمای سریع

    End-to-end text recognition in images using deep sequence modeling.

    یافته‌های کلیدی

    The proposed method achieved competitive and superior accuracy on standard benchmarks like IIIT-5K and ICDAR, while being end-to-end trainable and requiring no manual segmentation.

    هدف

    To create a unified trainable model that accurately recognizes variable-length text in natural images.

    روش

    The CRNN architecture combines a CNN for visual feature extraction, a bidirectional RNN (LSTM) for sequence modeling, and a CTC layer for mapping feature sequences to character outputs without requiring character-level segmentation.

    نتایج

    High accuracy on scene text benchmarks with simplified training and inference.

    نتیجه‌گیری

    CRNN is an effective and simple solution for text recognition and has become a foundational baseline for subsequent work.

    مفاهیم کلیدی

    OCR، text recognition، CRNN، CTC، LSTM

    مطالعه‌ی بیشتر

    https://ieeexplore.ieee.org/document/7649221

    تحلیل

    This paper has been highly influential in the OCR field due to its elegant architecture and strong performance, becoming a reference standard.

    محدودیت‌ها

    Performance degrades on very long texts, texts with large character spacing, and languages with special characters (e.g., Arabic), and the model requires large training datasets.

    کارهای آینده

    The authors suggested extending the model for multi-directional and multi-lingual text recognition, and incorporating attention mechanisms for further improvement.

    کاربرد عملی

    CRNN is used in automated registration systems, license plate reading, and document scanning. In the finger robot project, this model is ideal for recognizing text on the screen (button labels, window titles, text content) so the robot can read on-screen instructions and act accordingly (e.g., click 'Save' or type in a search field).

    References (this paper draws on) (0)

    No references recorded.

    Cited by (0) ▶

    No paper cites this one yet.

    Prerequisite learning path for this paper