Scholarlink

Zero-Shot Learning (ZSL) is a powerful learning paradigm that enables the classification of unseen classes by leveraging semantic information such as attributes or word embeddings. This paper presents a novel framework that utilizes a CAE to extract compact, discriminative, and semantically rich visual features directly from images. Unlike traditional approaches that rely on pre-trained deep networks like VGG16 or ResNet50, proposed framework employs unsupervised feature learning to reduce redundancy and dimensionality while preserving essential information. These learned visual features are mapped to semantic embeddings using an Artificial Neural Network (ANN) classifier. The model is evaluated on three benchmark datasets SUN, CUB, and AwA2 using semantic representations from FastText, BERT, and GloVe, with GloVe yielding the best results. Experimental results show a 5.7% improvement in classification accuracy over the Enhanced ZSL baseline framework, despite using approximately 50% reduction in feature dimensionality. The proposed framework not only enhances generalization to unseen classes but also achieves faster inference and lower computational cost, highlighting the effectiveness of CAE-based visual encoding in Zero-Shot image classification.

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

Autoencoder-Driven Visual Feature Learning for Zero-Shot Image Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ansari Shaista Khanam,
  • Poonam N. Sonar

摘要

Zero-Shot Learning (ZSL) is a powerful learning paradigm that enables the classification of unseen classes by leveraging semantic information such as attributes or word embeddings. This paper presents a novel framework that utilizes a CAE to extract compact, discriminative, and semantically rich visual features directly from images. Unlike traditional approaches that rely on pre-trained deep networks like VGG16 or ResNet50, proposed framework employs unsupervised feature learning to reduce redundancy and dimensionality while preserving essential information. These learned visual features are mapped to semantic embeddings using an Artificial Neural Network (ANN) classifier. The model is evaluated on three benchmark datasets SUN, CUB, and AwA2 using semantic representations from FastText, BERT, and GloVe, with GloVe yielding the best results. Experimental results show a 5.7% improvement in classification accuracy over the Enhanced ZSL baseline framework, despite using approximately 50% reduction in feature dimensionality. The proposed framework not only enhances generalization to unseen classes but also achieves faster inference and lower computational cost, highlighting the effectiveness of CAE-based visual encoding in Zero-Shot image classification.