Visual Saliency Prediction of Global Attention Based on Relevance Perception
Chenzhou Deng, Xiangyang Chen, Qianqian Cao · 2021
Target detection and tracking is a basic problem in the field of computer vision, and it is also the key and core technology of intelligent video surveillance systems. In order to effectively improve the saliency prediction task, this paper proposes an improved feature extraction method based on the residual neural network model. The architecture forms an encoder-decoder structure, and proposes an associated perception global attention module to capture multi-scale features in parallel. Through this module, global structural information can be captured for better attention learning, high-level visual feature extraction on multiple spatial scales, and global and local associated information for weight calculation, and the result is expressed as the global scene information. Combined to accurately predict visual saliency. The competitive and consistent results of multiple evaluation indicators are achieved on two public saliency benchmarks. Compared with the existing methods, it also has a certain lead compared with other saliency map prediction models.