Fine-grained classification based on multi-scale pyramid convolution networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fine-grained image classification is challenged by large intra-class and small inter-class variances.
• Existing methods often overlook multi-scale information, limiting their ability to detect subtle differences.

Purpose of the Study:

• To propose a novel weakly supervised fine-grained classification network leveraging multi-scale information.
• To enhance feature extraction capabilities for subtle variations in images.

Main Methods:

• Utilized a multi-scale pyramid with pyramid convolution kernels within a residual network to expand receptive fields.
• Incorporated a weakly supervised data augmentation network (WS-DAN) to mitigate overfitting.
• Introduced a novel attention module combining spatial and channel attention to focus on salient object parts.

Main Results:

• The proposed method demonstrated effectiveness in extracting subtle features crucial for classification.
• Experiments on three public benchmarks confirmed the network's ability to achieve accurate classification.

Conclusions:

• The multi-scale pyramid network effectively addresses limitations in current fine-grained classification approaches.
• The integration of WS-DAN and a novel attention module significantly improves model performance and robustness.