基于特征增强模型的广义零样本学习

Feature Enhancement Model for Generalized Zero-Shot Learning

摘要: 传统零样本学习旨在通过训练模型实现对未知类别的样本的精确分类，使模型具备对新任务和新环境的适应能力。但广义零样本学习的目标更为艰巨，它不仅要求模型能够辨识并分类未知类别的样本，还需确保对已知类别的样本也能准确无误地归入其对应的类别。由于在实际的训练过程中，我们仅能获得已知类别的样本，这使得零样本学习在分类任务中面临巨大的挑战。为了克服这一难题，我们创新性地提出了特征增强模型(Feature Enhancement Model，简称FE)。该模型不仅具备生成高质量未知类别样本的能力，以弥补训练样本的不足，而且能够构建每个样本的虚拟语义信息。此外，FE模型还配备了特征过滤模块，用于筛选出每个样本的核心特征。最终，模型将这些核心特征、样本本身以及虚拟语义信息相结合，作为最终的特征进行分类。这种方法通过凸显每类样本的独特性，有效地提升了分类的准确性和性能。

Abstract: Traditional zero-shot learning aims to achieve accurate classification of unseen class samples by training the model, so that the model can adapt to new tasks and new environments. However, the goal of generalized zero-shot learning is more difficult. It not only requires the model to identify and classify samples of unseen classes, but also needs to ensure that the samples of seen classes can be accurately classified into their corresponding classes. Because in the actual training process, we can only obtain samples of seen classes, which makes zero-shot learning in the classification task faces a huge challenge. To overcome this problem, we proposed an innovative feature enhancement model (FE). This model not only has the ability to generate high-quality unseen class samples to make up for the deficiency of training samples, but also can construct virtual semantic descriptions of each sample. In addition, the FE model is equipped with a feature filtering module to screen out the core features of each sample. Finally, the model combines these core features, the sample itself and the virtual semantic descriptions as the final features for classification. By highlighting the uniqueness of each type of sample, this method effectively improves the accuracy and performance of classification.