Fine-Grained Image Classification Using Modified DCNNs Trained by Cascaded Softmax and Generalized Large-Margin...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Fine-grained image classification presents challenges due to subtle inter-class variations and intra-class similarities.
• Deep Convolutional Neural Networks (DCNNs) are powerful tools but require specialized techniques for fine-grained tasks.
• Existing DCNN models often struggle to capture the hierarchical nature of fine-grained labels.

Purpose of the Study:

• To improve the accuracy of DCNN-based fine-grained image classification.
• To effectively model the hierarchical label structure inherent in fine-grained datasets.
• To develop a novel loss function that explicitly leverages hierarchical information and class similarities.

Main Methods:

• Introduced 'h' fully connected layers to replace the top layer of DCNNs, trained with cascaded softmax loss.
• Proposed a Generalized Large-Margin (GLM) loss function to exploit hierarchical label structures and similarity regularities.
• Designed a framework that is independent of specific DCNN architectures, allowing for broad applicability.

Main Results:

• Demonstrated significant improvements in fine-grained image classification accuracy across multiple DCNN models (AlexNet, GoogLeNet, VGG).
• Validated the effectiveness of the proposed methods on benchmark datasets including Stanford car, FGVC-Aircraft, and CUB-200-2011.
• Showcased the GLM loss's ability to reduce between-class similarity and within-class variance while enhancing intra-class similarity.

Conclusions:

• The proposed approach effectively enhances DCNN performance for fine-grained image classification.
• The integration of hierarchical modeling and the GLM loss function offers a robust solution for complex visual recognition tasks.
• The framework's independence from specific DCNN architectures makes it a versatile tool for advancing computer vision research.