Scholarlink

<p>Scene character recognition is essential for applications like traffic sign detection and license plate identification. However, it remains a challenging task due to the diversity in character shapes, colors, and fonts found in natural environments. In this paper, we introduce a novel hybrid classification approach that integrates deep features with machine learning classifiers through transfer learning. Specifically, we leverage pretrained CNN models including InceptionV3, VGG16, ResNet50V2, MobileNetV2, and DenseNet201, for deep feature extraction, while the K-nearest neighbors, the One-class Principal Component Classifier, and Support Vector Machines (SVM) are used for classification. Experiments are conducted on the ICDAR03-ch dataset, which comprises Latin scene character images. Given the limited sample size for each character, we apply data augmentation techniques using affine image transformations. The highest classification accuracy is achieved by the InceptionV3-SVM system, surpassing the current state of the art by 3%.</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

Learning deep features for scene Latin characters recognition

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Fatima Zouaoui,
  • Hassiba Nemmour,
  • Youcef Chibani

摘要

Scene character recognition is essential for applications like traffic sign detection and license plate identification. However, it remains a challenging task due to the diversity in character shapes, colors, and fonts found in natural environments. In this paper, we introduce a novel hybrid classification approach that integrates deep features with machine learning classifiers through transfer learning. Specifically, we leverage pretrained CNN models including InceptionV3, VGG16, ResNet50V2, MobileNetV2, and DenseNet201, for deep feature extraction, while the K-nearest neighbors, the One-class Principal Component Classifier, and Support Vector Machines (SVM) are used for classification. Experiments are conducted on the ICDAR03-ch dataset, which comprises Latin scene character images. Given the limited sample size for each character, we apply data augmentation techniques using affine image transformations. The highest classification accuracy is achieved by the InceptionV3-SVM system, surpassing the current state of the art by 3%.