Multi-scale semantic enhancement network for object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Feature Pyramid Networks (FPN) are crucial for extracting multi-scale information in object detection.
• Existing FPN methods face challenges with semantic gaps between features of different scales, leading to aliasing.
• This semantic gap hinders the effective fusion of multi-scale features.

Purpose of the Study:

• To propose a novel Multi-Scale Semantic Enhancement Feature Pyramid Network (MSE-FPN).
• To address the semantic gap and feature aliasing issues in FPN-based object detection.
• To enhance the overall performance of object detection models.

Main Methods:

• Introduced a Semantic Enhancement Module using self-attention to capture global context before feature fusion.
• Developed a Semantic Injection Module to integrate global semantic information across different feature scales.
• Implemented a Gated Channel Guidance Module to selectively filter features and mitigate aliasing during fusion.

Main Results:

• MSE-FPN integrated with Faster R-CNN achieved 39.4 AP (ResNet50) and 41.2 AP (ResNet101).
• Using ResNet-101-64x4d backbone, MSE-FPN reached 43.4 AP.
• Significant improvements in object detection performance were observed compared to standard FPN.

Conclusions:

• MSE-FPN effectively overcomes the limitations of traditional FPN architectures.
• The proposed modules enhance feature representation and reduce aliasing.
• MSE-FPN offers a superior alternative for state-of-the-art FPN-based object detectors.