HA-FPN: Hierarchical Attention Feature Pyramid Network for Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Object detection aims to identify and locate objects in images, with multi-scale processing enhancing accuracy.
• Existing methods often neglect contextual information between objects and may lose semantic details due to channel reduction for efficiency.
• Feature Pyramid Networks (FPNs) effectively extract multi-scaled features but can be improved by addressing contextual and semantic information loss.

Purpose of the Study:

• To enhance the accuracy and efficiency of object detection using attention mechanisms.
• To address the limitations of existing object detection models in capturing contextual information and preserving semantic details.
• To propose a novel Hierarchical Attention Feature Pyramid Network (HA-FPN) for improved object detection performance.

Main Methods:

• Proposed a novel Hierarchical Attention Feature Pyramid Network (HA-FPN).
• Incorporated Transformer Feature Pyramid Networks (TFPNs) to capture intra- and inter-scale contextual information using self-attention on embedded features.
• Utilized Channel Attention Modules (CAMs) to select informative channels and mitigate semantic information loss.

Main Results:

• The HA-FPN significantly improves bounding box detection accuracy.
• The model achieves more precise identification and localization of target objects.
• Experiments on the MS COCO dataset show superior performance compared to existing multi-object detection models with minimal computational overhead.

Conclusions:

• The proposed HA-FPN effectively integrates contextual information and attention mechanisms to boost object detection performance.
• The method enhances feature representational power and efficiency, leading to state-of-the-art results.
• HA-FPN offers a promising approach for accurate and efficient object detection in complex visual scenes.