Global and Local Channel-based Energy-Edge Attention Mechanism in Deep Network for Visual Saliency Prediction
S. Caroline, Y. Jacob Vetha Raj · 2025
Visual saliency prediction is the field of computer vision that helps to identify the salient region of the image which has numerous applications in scene understanding. This paper proposes a deep convolutional network with Global and Local channel-based energy edge attention mechanisms (GLEE) for Visual Saliency Prediction. The approach uses four different attention modules namely global channel-energy attention (GEN), local channel-energy attention (LEN), global channeledge attention (GED), and local channel-edge attention (LED). The GEN and GED attention are used to preserve the energy and edge components of the contextual features between the channels. In contrast, the LEN and LED attention are used to preserve the energy and edge components of the descriptor of a channel. The global attention map is estimated from the single energy/edge map estimated from the feature map of the convolutional filter, while the local attention map is estimated from the energy/edge map estimated from each convolutional filter. The evaluation of the proposed visual saliency prediction approach was performed using scales such as SIM, KL, NSS, CC, AUC, and SAUC using the dataset MIT-300 and SALICON. In the case of the MIT-300 dataset the proposed GLEE results in KL, SIM, and NSS of 0.392, 0.696, and 0.811 respectively when evaluated using the MIT-300 dataset. The approach also yields an AUC and SAUC of 0.873 and 0.753 respectively when evaluated using the SALICON dataset. The results illustrate that the proposed GLEE approach performs better than other similar visual saliency prediction schemes.