CORE-Net: A cross-modal orthogonal representation enhancement network for low-altitude multispectral object detection

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Object detection in visible light (RGB) is limited by low illumination.
• Infrared (IR) imaging offers robustness in low-light but requires fusion with RGB.
• Existing multispectral fusion methods are computationally intensive and complex.

Purpose of the Study:

• To develop an efficient and high-performance multispectral object detection network.
• To reduce the computational overhead and architectural complexity of fusion methods.
• To enhance object detection accuracy in challenging illumination conditions.

Main Methods:

• Proposed a novel Cross-modal Orthogonal Representation Enhancement Network (CORE-Net).
• Employed a dual-branch architecture with a Cross-modal Concatenation Network Framework (CCNF) for efficient feature integration.
• Introduced Multiple Pooling Convolution Downsampling (MPCD) and Refined Integration Network (RINet) modules for optimized feature extraction.

Main Results:

• CORE-Net achieved state-of-the-art (SOTA) performance on DroneVehicle and LLVIP datasets.
• Demonstrated superior detection accuracy and computational efficiency compared to existing methods.
• Ablation studies confirmed the effectiveness of MPCD and RInet components.

Conclusions:

• CORE-Net offers an efficient and robust paradigm for multispectral object detection.
• The model effectively suppresses background noise and enhances fine-grained features.
• Practical deployment on edge devices validates its efficiency.