Scholarlink

<p>Fine-Grained Visual Classification (FGVC) is a challenging task due to the small inter-class differences and large intra-class variations inherent in datasets. These characteristics pose significant challenges to traditional image classification models. Existing FGVC methods based on contrastive learning typically generate positive and negative samples using images from the same or different classes. However, they often overlook the substantial redundant information within a single image, which can impair classification performance. Moreover, the construction of multi-sample contrastive learning strategies inevitably increases the training burden on the model. To address these issues, this paper proposes a fine-grained visual classification method based on contrastive learning regularization. The proposed approach focuses on constructing a contrastive learning process from a single image to enhance discriminative feature representation, thereby eliminating the reliance on additional image data of existing contrastive learning-based methods and fully exploiting useful information within the image itself. Specifically, we first design a contrastive interaction attention module, which leverages contrastive learning strategies to guide spatial attention toward different mutually exclusive regions, thereby producing more precise and discriminative attention distributions. Subsequently, a feature boosting module is designed to augment the representation of discriminative features and extract complementary information between feature representations, thus improving feature distinctiveness. This module also enables the model to capture rich semantic information from multi-scale features, enhancing its adaptability to objects at varying scales. Furthermore, to ensure comprehensive representation of fine-grained features, we incorporate a knowledge distillation framework, allowing the network to achieve higher classification accuracy with faster convergence. Experimental results demonstrate that the proposed method exhibits strong competitiveness across multiple evaluation metrics on three publicly available fine-grained datasets as well as on a self-constructed algae dataset, validating the effectiveness of our approach. The project code is openly accessible at <a href="https://github.com/ShuoWang00/SGC-main">https://github.com/ShuoWang00/SGC-main</a>.</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

SGC-Net: an effective fine-grained visual classification framework via self-guided contrast regularization and feature boosting

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Shuo Wang,
  • Bohao Wen,
  • Xueqing Zhang,
  • Jin Xiong,
  • Jianlei Liu,
  • Fengjuan Feng

摘要

Fine-Grained Visual Classification (FGVC) is a challenging task due to the small inter-class differences and large intra-class variations inherent in datasets. These characteristics pose significant challenges to traditional image classification models. Existing FGVC methods based on contrastive learning typically generate positive and negative samples using images from the same or different classes. However, they often overlook the substantial redundant information within a single image, which can impair classification performance. Moreover, the construction of multi-sample contrastive learning strategies inevitably increases the training burden on the model. To address these issues, this paper proposes a fine-grained visual classification method based on contrastive learning regularization. The proposed approach focuses on constructing a contrastive learning process from a single image to enhance discriminative feature representation, thereby eliminating the reliance on additional image data of existing contrastive learning-based methods and fully exploiting useful information within the image itself. Specifically, we first design a contrastive interaction attention module, which leverages contrastive learning strategies to guide spatial attention toward different mutually exclusive regions, thereby producing more precise and discriminative attention distributions. Subsequently, a feature boosting module is designed to augment the representation of discriminative features and extract complementary information between feature representations, thus improving feature distinctiveness. This module also enables the model to capture rich semantic information from multi-scale features, enhancing its adaptability to objects at varying scales. Furthermore, to ensure comprehensive representation of fine-grained features, we incorporate a knowledge distillation framework, allowing the network to achieve higher classification accuracy with faster convergence. Experimental results demonstrate that the proposed method exhibits strong competitiveness across multiple evaluation metrics on three publicly available fine-grained datasets as well as on a self-constructed algae dataset, validating the effectiveness of our approach. The project code is openly accessible at https://github.com/ShuoWang00/SGC-main.