DSPNet: Dual-Path Saliency Prediction With Reconstructed Feature Injection and Discrepancy-Aware KL Loss

Visual saliency prediction (VSP) is a fundamental task in computer vision, which enables efficient sensing and computation by prioritizing informative regions, such as object detection, scene understanding, and video compression. However, the VSP task still faces three challenges: (i) fine-grained texture is irreversibly lost during repeated down-sampling in single-stream convolutional networks; (ii) traditional attention modules apply fixed, simplistic weights that adapt poorly to complex scenes; and (iii) standard Kullback-Leibler (KL) divergence fails to sufficiently amplify gradients in weakly activated regions, thereby hindering the learning of hard samples. A dual-path saliency prediction network (DSPNet) is proposed to address the aforementioned challenges. A vector-quantized variational autoencoder (VQ-VAE) branch discretizes local textures and reinjects them into a U-Net backbone to recover lost details caused by down-sampling. The restored texture is integrated with high-level semantics via a bidirectional mutual transpose attention (BMTA) decoder, enabling nonlinear cross-layer interactions and enhancing spatial discrimination in complex scenes. A discrepancy-aware focal KL (F-KL) loss is designed to optimize the model training, which adaptively emphasizes discrepant regions to improve attention to areas with prediction gaps. Evaluations on four public benchmarks (MIT300, SALICON, PASCAL-S, and TORONTO) show state-of-the-art results across metrics such as Similarity (SIM), KL divergence, and Normalized Scanpath Saliency (NSS), confirming that DSPNet effectively balances local texture fidelity with global semantic understanding.