Scholarlink

Purpose: <p><i>ChitraKavya</i> is a fusion of poetry and art, where the poetic text flows within artistic images having architectural forms, geometric shapes, or depictions of living and non-living objects. The characters embedded in complex patterns pose major challenges for automated text recognition. This study focuses on Odia <i>ChitraKavya</i>s and aims to develop a robust method for spotting and recognizing text in such complex artistic layouts.</p> Methods: <p>We propose a novel <i>ChitraKavya Text Spotting and Recognition Framework (CTSRF)</i>, a unified hybrid framework that integrates dataset augmentation, character spotting, and character recognition within a single pipeline. CTSRF combines a Generative Adversarial Network (GAN) based augmentation module with a multi-stage, script-aware character spotting pipeline based on Connected Component Analysis (CCA), Maximally Stable Extremal Regions (MSER), Non-Maximum Suppression (NMS), and Intersection over Union (IoU) refinement. As part of CTSRF, we introduce <i>OdiaOCRNet</i>, a dedicated character recognition model designed for Odia scripts, built on a ConvNeXt backbone and augmented with a Convolutional Block Attention Module (CBAM) to enhance discriminative feature learning for complex Odia character forms. The proposed framework is evaluated against state-of-the-art scene-text detectors and object-detection-based baselines, including YOLO, which are shown to be inadequate for the highly curved and ornamentally embedded text layouts characteristic of <i>ChitraKavyas</i>.</p> Results: <p>On Odia <i>ChitraKavya</i> images, the proposed CTSRF framework achieves up to 35% higher character-level spotting accuracy than Google Vision, Segment Anything Model, and CRAFT, and improves OCR accuracy by up to 55% over Google Vision. It also outperforms fine-tuned Vision Transformer (ViT-B16), Data-efficient Image Transformer (DeiT), and Class-Attention Image Transformer (CaiT) models by up to 4.5%.</p> Conclusions: <p>The proposed hybrid framework significantly improves text spotting and recognition in Odia <i>ChitraKavya</i>s. The approach establishes a foundation for OCR in pictorial poem manuscripts and can be extended to other artistic and non-linear text domains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Method for Spotting and Recognizing Texts in Odia ChitraKavyas

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Swarupananda Bissoyi,
  • Santosh Kumar Das,
  • Siba Kumar Udgata

摘要

Purpose:

ChitraKavya is a fusion of poetry and art, where the poetic text flows within artistic images having architectural forms, geometric shapes, or depictions of living and non-living objects. The characters embedded in complex patterns pose major challenges for automated text recognition. This study focuses on Odia ChitraKavyas and aims to develop a robust method for spotting and recognizing text in such complex artistic layouts.

Methods:

We propose a novel ChitraKavya Text Spotting and Recognition Framework (CTSRF), a unified hybrid framework that integrates dataset augmentation, character spotting, and character recognition within a single pipeline. CTSRF combines a Generative Adversarial Network (GAN) based augmentation module with a multi-stage, script-aware character spotting pipeline based on Connected Component Analysis (CCA), Maximally Stable Extremal Regions (MSER), Non-Maximum Suppression (NMS), and Intersection over Union (IoU) refinement. As part of CTSRF, we introduce OdiaOCRNet, a dedicated character recognition model designed for Odia scripts, built on a ConvNeXt backbone and augmented with a Convolutional Block Attention Module (CBAM) to enhance discriminative feature learning for complex Odia character forms. The proposed framework is evaluated against state-of-the-art scene-text detectors and object-detection-based baselines, including YOLO, which are shown to be inadequate for the highly curved and ornamentally embedded text layouts characteristic of ChitraKavyas.

Results:

On Odia ChitraKavya images, the proposed CTSRF framework achieves up to 35% higher character-level spotting accuracy than Google Vision, Segment Anything Model, and CRAFT, and improves OCR accuracy by up to 55% over Google Vision. It also outperforms fine-tuned Vision Transformer (ViT-B16), Data-efficient Image Transformer (DeiT), and Class-Attention Image Transformer (CaiT) models by up to 4.5%.

Conclusions:

The proposed hybrid framework significantly improves text spotting and recognition in Odia ChitraKavyas. The approach establishes a foundation for OCR in pictorial poem manuscripts and can be extended to other artistic and non-linear text domains.