A Fine-Grained Image Classification and Detection Method Based on Convolutional Neural Network Fused with Attention...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Fine-grained image classification is challenging due to subtle inter-class differences and intra-class variations.
• Attention mechanisms are crucial for focusing on discriminative regions in images, aiding classification.
• Complex backgrounds can interfere with accurate image recognition.

Purpose of the Study:

• To propose a novel fine-grained image classification method, WSFF-BCNN, enhancing discriminant regional feature extraction.
• To improve the loss function and feature extraction process within convolutional neural networks for classification tasks.
• To effectively fuse multi-scale features for a richer image representation.

Main Methods:

• Employs a mixed attention mechanism (channel and spatial domains) to highlight key feature map regions.
• Utilizes an improved bilinear model with ResNet50 backbone to extract and fuse multi-scale features.
• Incorporates weak supervision feature fusion and an enhanced loss function during network training.

Main Results:

• The WSFF-BCNN method effectively extracts detailed descriptive information by focusing on channel and spatial attention.
• Multi-scale features from the improved bilinear model capture both spatial location and low-level image characteristics.
• Feature fusion via bilinear pooling generates a comprehensive image feature representation.

Conclusions:

• The proposed WSFF-BCNN method enhances fine-grained image classification accuracy by leveraging weak supervision and advanced feature fusion techniques.
• Mixed attention and multi-scale feature extraction are key components for addressing the challenges in classifying visually similar objects.
• This approach offers a robust solution for distinguishing subtle differences critical in fine-grained visual categorization.