unir/MastersThesis
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Leyenda
Some checks failed
build_docker / essential (push) Successful in 1s
Details
build_docker / build_paddle_ocr (push) Successful in 4m0s
Details
build_docker / build_paddle_ocr_gpu (push) Successful in 18m53s
Details
build_docker / build_easyocr (push) Successful in 16m12s
Details
build_docker / build_easyocr_gpu (push) Successful in 22m37s
Details
build_docker / build_doctr (push) Successful in 21m22s
Details
build_docker / build_raytune (push) Successful in 2m50s
Details
build_docker / build_doctr_gpu (push) Has been cancelled
Details

Browse Source
This commit is contained in:
Sergio Jimenez Jimenez
2026-02-04 19:56:30 +01:00
parent d746a3c73f
commit 868f748a8d
4 changed files with 617 additions and 357 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
docs/00_resumen.md
View File

@@ -6,7 +6,7 @@ Se realizó un estudio comparativo de tres soluciones OCR de código abierto: Ea
Los resultados demuestran que la optimización de hiperparámetros logró mejoras significativas: el mejor trial individual alcanzó un CER de 0.79% (precisión del 99.21%), cumpliendo el objetivo de CER < 2%. Al validar la configuración optimizada sobre el dataset completo de 45 páginas, se obtuvo una mejora del 12.8% en CER (de 8.85% a 7.72%). El hallazgo más relevante fue que el parámetro `textline_orientation` (clasificación de orientación de línea de texto) tiene un impacto crítico en el rendimiento. Adicionalmente, se identificó que el umbral de detección (`text_det_thresh`) presenta una correlación positiva moderada (0.43) con el error, lo que indica que valores más bajos tienden a mejorar el rendimiento.
**Fuente:** [`docs/metrics/metrics_paddle.md`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/docs/metrics/metrics_paddle.md), [`src/results/correlations/paddle_correlations.csv`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/src/results/correlations/paddle_correlations.csv).
**Fuente:** [`metrics_paddle.md`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/docs/metrics/metrics_paddle.md), [`paddle_correlations.csv`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/src/results/correlations/paddle_correlations.csv).
Este trabajo demuestra que la optimización de hiperparámetros es una alternativa viable al fine-tuning, especialmente útil cuando se dispone de modelos preentrenados para el idioma objetivo. La infraestructura dockerizada desarrollada permite reproducir los experimentos y facilita la evaluación sistemática de configuraciones OCR.
@@ -22,7 +22,7 @@ A comparative study of three open-source OCR solutions was conducted with EasyOC
Results demonstrate that hyperparameter optimization achieved significant improvements. The best individual trial reached a CER of 0.79% (99.21% accuracy), meeting the CER < 2% objective. When validating the optimized configuration on the full 45-page dataset, a 12.8% CER improvement was obtained (from 8.85% to 7.72%). The most relevant finding was that the `textline_orientation` parameter (text line orientation classification) has a critical impact on performance. Additionally, the detection threshold (`text_det_thresh`) showed a moderate positive correlation (0.43) with error, indicating that lower values tend to improve performance.
Sources: [`docs/metrics/metrics_paddle.md`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/docs/metrics/metrics_paddle.md), [`src/results/correlations/paddle_correlations.csv`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/src/results/correlations/paddle_correlations.csv).
Sources: [`metrics_paddle.md`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/docs/metrics/metrics_paddle.md), [`paddle_correlations.csv`](https://seryus.ddns.net/unir/MastersThesis/src/branch/main/src/results/correlations/paddle_correlations.csv).
This work demonstrates that hyperparameter optimization is a viable alternative to fine-tuning, especially useful when pre-trained models for the target language are available. The dockerized infrastructure developed enables experiment reproducibility and facilitates systematic evaluation of OCR configurations.