PROJECT TITLE :

RGB-T Salient Object Detection via Fusing Multi-Level CNN Features

ABSTRACT:

Deep convolutional Neural Networks have lately made significant progress in the field of RGB-inducing salient object recognition (CNNs). The problem is that these detections have to contend with a wide range of tough situations that include low-light circumstances, varying light levels, and crowded backdrops. Instead of trying to improve RGB-based saliency detection, this research takes advantage of the complementing advantages of RGB and thermal infrared images. A novel end-to-end network for multi-modal salient object recognition is proposed, which turns the challenge of RGB-T saliency detection into a CNN feature-fusion issue. An initial backbone network is used to extract coarse features from each RGB or thermal infrared image, and then several ADFC modules are designed to extract multi-level refined features for each single-modal input image, taking into account that features captured at different depths differ in semantic information and visual details. Additionally, the cross-modal features from a pair of RGB-T images are combined using an MGF module, which utilises the ADFC modules for each level. As a final step, a joint attention guided bi-directional message passing (JABMP) module integrates the multi-level fused features from MGF modules to forecast saliency. Public RGB-T salient object datasets show that our proposed algorithm outperforms current state-of-the art algorithms in tough settings such as low contrast lighting and complex background conditions.