ResneSt-Transformer: Joint attention segmentation-free for end-to-end handwriting paragraph recognition model
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Offline handwritten text recognition (HTR) typically relies on segmented text-line images for training and transcription. However, acquiring line-level position and transcript information can be challenging and time-consuming, while automatic line segmentation algorithms are prone to errors that impede the recognition phase. To address these issues, we introduce a state-of-the-art solution that integrates vision and language models using efficient split and multi-head attention neural networks, referred to as joint attention (ResneSt-Transformer), for end-to-end recognition of handwritten paragraphs. Our proposed novel one-stage, segmentation-free pipeline employs joint attention mechanisms to process paragraph images in an end-to-end trainable manner. This pipeline comprises three modules, with the output of one serving as the input for the next. Initially, a feature extraction module employing a CNN with a split attention mechanism (ResneSt50) is utilized. Subsequently, we develop an encoder module containing four transformer layers to generate robust representations of the entire paragraph image. Lastly, we designed a decoder module with six transformer layers to construct weighted masks. The encoder and decoder modules incorporate a multi-head self-attention mechanism and positional encoding, enabling the model to concentrate on specific feature maps at the current time step. By leveraging joint attention and a segmentation-free approach, our neural network calculates split attention weights on the visual representation, facilitating implicit line segmentation. This strategy signifies a substantial advancement toward achieving end-to-end transcription of entire paragraphs. Experiments conducted on paragraph-level benchmark datasets, including RIMES, IAM, and READ 2016 test datasets, demonstrate competitive results compared to recent paragraph-level models while maintaining reduced complexity. The code and pre-trained models are available on our GitHub repository here: HTTPS link.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it