Adaptive decoupling-fusion in Siamese network for image classification.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel at semantic understanding but often lose visual details.
• Existing hierarchical models struggle to retain fine-grained appearance information.
• Traditional Siamese Networks with shared weights limit feature diversity.

Purpose of the Study:

• To introduce an Adaptive Decoupling Fusion (ADF) method for preserving visual details in CNNs.
• To enhance semantic understanding by integrating shallow-layer appearance information into deep features.
• To improve the adaptability and performance of deep learning models for visual tasks.

Main Methods:

• Developed an Adaptive Decoupling Fusion (ADF) approach using a Siamese Network architecture.
• Decoupled appearance information from one branch and embedded it into the deep feature space of another.
• Implemented differentiated collaborative learning with distinct weights for network branches.
• Introduced a Mapper module with depthwise separable convolution and gated fusion for information flow regulation.

Main Results:

• Achieved a top-1 accuracy of 81.11% on the ImageNet-1k dataset.
• Demonstrated effectiveness under fully self-supervised conditions with minimal data augmentation.
• Showcased the ability of ADF-ResNeXt-101 to preserve and leverage visual details for improved performance.

Conclusions:

• The proposed Adaptive Decoupling Fusion (ADF) effectively preserves crucial visual details lost in standard CNNs.
• ADF enhances semantic understanding by synergistically combining appearance and semantic information.
• This method offers a promising direction for improving deep learning models in computer vision tasks.