Significant feature suppression and cross-feature fusion networks for fine-grained visual classification.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Fine-grained visual classification (FGVC) benefits from extracting distinguishing features from object parts.
• Attention mechanisms in computer vision focus on discriminative regions but neglect less prominent ones and may not fully explore feature hierarchies.
• Existing methods often fail to leverage the intrinsic connections between higher-order and lower-order features for optimal classification.

Purpose of the Study:

• To propose a novel network, SFSCF-Net, for enhanced fine-grained visual classification.
• To address limitations of current attention mechanisms by exploring interactions between higher-order feature representations.
• To improve classification performance by integrating saliency feature suppression and cross-feature fusion.

Main Methods:

• An object-level image generator (OIG) creates object masks to reduce background interference.
• A saliency feature suppression module (SFSM) accurately masks the most distinguishing object parts.
• A cross-feature fusion method (CFM) interactively integrates features from different network layers for enriched semantic information.

Main Results:

• The proposed SFSCF-Net model was trained end-to-end.
• The model achieved state-of-the-art or competitive results on four benchmark FGVC datasets.
• The integrated higher-order features significantly contributed to the model's classification decision-making.

Conclusions:

• SFSCF-Net effectively explores interaction learning between different higher-order feature representations.
• The model demonstrates superior performance in fine-grained visual classification tasks.
• The proposed methods enhance feature utilization and semantic richness for improved accuracy.