Lightweight transformer-driven multi-scale trapezoidal attention network for saliency detection

Muhammad Talha Usman, Habib Khan, Imad Rida, Jakeoung Koo

Abstract: Highlights•The proposed network balances efficiency and performance via a lightweight design.•CFRBs with dilated convolutions enhance multi-scale contextual representations.•The introduced TAM effectively refines features by integrating ASCA and CCG.•FAMHA blocks capture long-range dependencies and global cues.•TRSNet outperforms SOTA methods across six saliency detection benchmarks.