A salient object detection model based on feature attention refinement

In recent years, substantial advancements have occurred in salient object detection algorithms, where the selection and appropriate integration of multi-scale features are crucial. A salient object recognition model utilising feature attention refinement is presented to mitigate information redundancy in current feature integration methods. A Semantic Attention Propagation Module (SAPM) is initially utilised in the decoder to analyse deep features imbued with semantic information using an attention mechanism, thereby acquiring global contextual knowledge. This information is subsequently input into each layer of the decoder via the globally directed flow for supervised training. A Contextual Feature Integration Module (CFIM) is employed to efficiently combine the multi-scale features extracted by the encoder with global contextual information, followed by refinement in the Layer-wise Feature Refinement Module (LFRM) to produce clear and comprehensive saliency maps. Experiments conducted on five public datasets demonstrate that AstraNet achieves competitive performance against the evaluated salient object detection methods while maintaining real-time efficiency. For input images of dimensions 320×320, the complete AstraNet achieves an inference speed of 35 frames per second (FPS).