WaveNet: Wavelet Network With Knowledge Distillation for RGB-T Salient Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Traditional convolutional neural networks (CNNs) and transformers face limitations in efficiency and generalization for computer vision tasks.
• Multilayer perceptrons (MLPs) offer reduced inductive bias and improved generalization compared to CNNs and transformers.
• Transformers, while powerful, incur significant increases in training and inference times.

Purpose of the Study:

• To introduce WaveNet, a novel architecture for salient object detection in RGB-thermal infrared images.
• To leverage wavelet-based MLPs for efficient and effective feature extraction.
• To enhance cross-modality fusion and learning through knowledge distillation and regularization techniques.

Main Methods:

• Proposed WaveNet architecture utilizing a vision task-oriented wavelet-based MLP for feature extraction.
• Employed knowledge distillation from a transformer teacher network to guide WaveNet learning.
• Utilized Kullback-Leibler distance for RGB and thermal infrared feature alignment.
• Implemented a progressively cascaded sine-cosine module for cross-layer feature fusion.

Main Results:

• WaveNet demonstrated impressive performance on benchmark RGB-thermal infrared datasets.
• The wavelet-based MLP effectively extracted salient features for object detection.
• Cross-modality fusion using discrete wavelet transform and the sine-cosine module proved successful.

Conclusions:

• WaveNet offers a promising approach for salient object detection in multi-modal imagery.
• The proposed architecture balances efficiency and performance, overcoming limitations of existing models.
• The integration of wavelet transforms and MLP offers a novel direction for computer vision research.