An Image Classification Method Based on Adaptive Attention Mechanism and Feature Extraction Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) offer high fault tolerance and computing capacity for image classification.
• Deeper CNNs improve fitting ability but can lead to higher training errors and reduced performance.
• Existing CNN architectures face challenges in balancing network depth and classification accuracy.

Purpose of the Study:

• To address the limitations of traditional CNNs in image classification.
• To propose a novel feature extraction network, AA-ResNet, with an adaptive attention mechanism.
• To enhance the feature representation ability and reduce overfitting in deep neural networks.

Main Methods:

• Developed AA-ResNet, incorporating an adaptive attention mechanism within residual modules.
• Utilized a pattern-guided feature extraction network to capture multi-level image features.
• Employed a pre-trained generator and a complementary network for enhanced feature learning.
• Trained the model using a multitask loss function with specialized classification for easily confused categories.

Main Results:

• AA-ResNet demonstrated strong performance in image classification tasks.
• The model achieved high fitting speed and accuracy on benchmark datasets (Cifar-10, Caltech-101, Caltech-256).
• The adaptive attention mechanism effectively utilized both global and local image information, enhancing feature representation.

Conclusions:

• AA-ResNet offers a promising solution for improving CNN-based image classification.
• The proposed adaptive attention mechanism effectively mitigates overfitting and enhances model accuracy.
• The method shows significant potential for various image classification applications requiring high accuracy and efficiency.