Decision Fusion Networks for Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are prevalent for image classification, aggregating features layer by layer.
• Information loss can occur during feature aggregation and abstraction in traditional CNNs.
• Lower-layer features may not be fully utilized in deep network architectures.

Purpose of the Study:

• To introduce a novel decision fusion module (DFM) to mitigate information loss in CNNs.
• To enhance feature aggregation and abstraction by incorporating intermediate decisions.
• To improve the performance of image classification networks.

Main Methods:

• Proposed a decision fusion module (DFM) integrated into CNNs.
• DFM makes intermediate decisions on auxiliary categories to guide feature learning.
• Fused intermediate decisions with original features before passing to subsequent layers.
• Stacked DFMs to create a decision fusion network for progressive feature refinement.

Main Results:

• DFM significantly improved performance across various common classification networks.
• Achieved superior results compared to state-of-the-art decision fusion methods.
• Demonstrated minimal additional computational cost.
• Validated generalization to object detection and semantic segmentation tasks.

Conclusions:

• The proposed DFM effectively reduces information loss and enhances feature discriminability in CNNs.
• DFM offers a computationally efficient method for improving deep learning model performance.
• The approach shows promise for broader applications in computer vision beyond image classification.