Related Experiment Video
Updated: Jul 28, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
Adaptive cognitive driven cross modal network for few shot fine grained recognition
1Faculty of Science and Technology, Beijing Normal-Hong Kong Baptist University, Zhuhai, 519087, China. t330026052@mail.bnbu.edu.cn.
Scientific Reports
|July 15, 2026
Summary
Adaptive cognitive driven cross modal network (ACD-Net) improves few-shot fine-grained recognition by addressing texture loss and knowledge integration. ACD-Net achieves state-of-the-art results, enhancing accuracy in challenging recognition tasks.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Few-shot learning (FSL) faces challenges in fine-grained recognition, including texture loss, poor cross-modal knowledge integration, and catastrophic forgetting.
- Existing FSL methods often fail to capture subtle textural details and effectively transfer knowledge from different modalities.
Purpose of the Study:
- To introduce a novel Adaptive cognitive driven cross modal network (ACD-Net) designed to overcome the limitations of current FSL methods for fine-grained recognition.
- To enhance the robustness and accuracy of few-shot learning models by integrating cognitive principles and advanced deep learning techniques.
Main Methods:
- The proposed ACD-Net incorporates three key innovations: Adaptive Dual-Domain Cognitive Attention (ADCA) for texture decoupling and region localization, Graph-Guided Semantic-to-Visual Distillation (GSD) for embedding semantic priors, and Dynamic Balanced Anti-forgetting (DBAF) loss to mitigate catastrophic forgetting.
- ADCA utilizes Two-Dimensional Discrete Wavelet Transform and Gated Recurrent Units; GSD employs Graph Convolutional Networks and bilinear attention; DBAF adaptively adjusts regularization weights.
Main Results:
- ACD-Net achieved state-of-the-art performance on benchmark datasets including miniImageNet, CUB-200-2011, and Medical-44.
- The model demonstrated significant improvements, elevating average accuracy by 1.75% in 1-shot and 1.36% in 5-shot learning scenarios.
Conclusions:
- ACD-Net presents a groundbreaking approach for few-shot fine-grained recognition, effectively addressing texture loss, cross-modal integration, and catastrophic forgetting.
- The network offers practical solutions for real-world applications like industrial defect inspection and clinical diagnosis.