Wavelet-Driven Multi-Band Feature Fusion for RGB-T Salient Object Detection.

Area of Science:

• Computer Vision
• Image Processing

Background:

• RGB-T salient object detection (SOD) is crucial in computer vision.
• Existing SOD methods struggle with integrating high- and low-frequency features across scales.
• This limitation hinders optimal detection performance in complex scenarios.

Purpose of the Study:

• To propose an advanced RGB-T salient object detection method.
• To enhance feature utilization by integrating wavelet transform and channel-wise attention fusion.
• To improve the detection of both global context and fine-grained details.

Main Methods:

• Utilized wavelet transform for feature differentiation and extraction of spatial characteristics.
• Employed a channel-wise criss-cross module (CCM) for adaptive cross-modal feature fusion.
• Integrated a feature selection wavelet transform module (FSW) for selecting beneficial low- and high-frequency features.

Main Results:

• The proposed method effectively extracts spatial characteristics, improving detection of global context and fine-grained details.
• Channel-wise attention fusion adaptively adjusts feature importance, generating rich fusion information.
• The FSW module enhances feature aggregation through long-distance connections, leading to higher segmentation accuracy.

Conclusions:

• The developed RGB-T SOD method significantly outperforms 22 state-of-the-art approaches.
• Wavelet transform and channel-wise attention fusion are effective in addressing limitations of existing SOD methods.
• The approach demonstrates superior performance in salient object detection tasks.