On the Imaginary Wings: Text-Assisted Complex-Valued Fusion Network for Fine-Grained Visual Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Fine-grained visual classification (FGVC) faces challenges due to high similarity between classes and significant variation within classes.
• Existing methods struggle to effectively model both data uncertainty and inter-class relationships simultaneously.

Purpose of the Study:

• To propose a novel approach, the text-assisted complex-valued fusion network (TA-CFN), for enhanced FGVC.
• To leverage complex values for modeling data uncertainty and graph convolutional networks (GCNs) for learning inter-class dependencies.

Main Methods:

• Features are expanded from 1-D real values to 2-D complex values, extending deep convolutional neural networks to the complex domain.
• Complex features are fused using complex projection and modulus operations.
• An undirected graph is constructed over object labels using a text corpus, and a GCN maps this graph into classifiers.

Main Results:

• Complex features provide a richer algebraic structure to model intra-class variations.
• GCNs capture inter-class dependencies, addressing subtle variations between categories.
• The TA-CFN achieves state-of-the-art performance on two widely used FGVC datasets.

Conclusions:

• The proposed TA-CFN effectively addresses the core challenges in FGVC by integrating complex-valued features and GCNs.
• This approach offers a promising direction for improving the accuracy and robustness of fine-grained visual classification models.