Weakly Alignment-Free RGBT Salient Object Detection With Deep Correlation Network.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• RGB-Thermal Salient Object Detection (SOD) typically requires well-aligned image pairs.
• Acquiring aligned RGBT image pairs is labor-intensive and costly.
• Existing methods struggle with unaligned RGBT data, limiting practical applications.

Purpose of the Study:

• To develop a novel deep learning model for weakly alignment-free RGBT SOD.
• To address the challenge of unaligned visible and thermal infrared image pairs.
• To improve the accuracy and efficiency of salient object detection in RGBT data.

Main Methods:

• Proposed a Deep Correlation Network (DCNet) for weakly alignment-free RGBT SOD.
• Introduced a modality alignment module using spatial affine, feature-wise affine transformations, and dynamic convolution.
• Developed a bi-directional decoder with a modality correlation ConvLSTM for hierarchical feature decoding.

Main Results:

• DCNet demonstrated remarkable performance on three public benchmark datasets.
• The proposed method significantly outperformed existing state-of-the-art RGBT SOD approaches.
• The modality alignment and bi-directional decoder effectively handled unaligned image pairs.

Conclusions:

• The proposed DCNet offers a robust solution for alignment-free RGBT SOD.
• The method effectively models cross-modal correlations for improved salient object detection.
• This work advances the practical applicability of RGBT SOD by reducing alignment requirements.