The Research of Salient Object Detection on RGB-T Multi-Source Image
Lixin An, Jikun Chen · 2021 IEEE 21st International Conference on Communication Technology (ICCT) · 2021
The images captured from visible light is limited by environment and illumination, which would make the result of salient object detection inaccurate. To solve the problem caused by this inherent property, thermal infrared image is introduced. Thermal infrared-captured images are able to be assistant of visible light images in detection due to their different imaging mechanism and factors of limitations, which make them complement each other. The paper proposes to combine the multisource image data of visible light and thermal infrared light images, and use the dual-channel convolutional neural network to fully extract the features of the two modalities. Additionally, the attention mechanism can prompt the neural network focus on essential parts, attention mechanism is introduced in the feature extraction part for improving accuracy.