Holistic Features and Deep-Guided Depth-Induced Mutual-Attention-Based Complex Salient Object Detection

Abstract

The multi-stream-based convolution neural network is the recent trend in saliency computation and receiving tremendous research interest. The existing models preferred the color- or depth-modality-based independent stream to extract saliency features, which are insufficient in complex and challenging scenarios. The proposed model produces an enormous set of saliency features (holistic feature space) by three-stream networks to target the above challenges. Two streams are independent, and the third one depends on the other to acquire all essential features. These holistic-feature-based learning include non-complementary, cross-complementary, and intra-complementary features. These features are utilized in middle-level fusion strategies to produce complete and correct salient objects. A novel, deeply guided depth-induced mutual attention map is proposed to start the cross- and intra-complementary fusion. This deep learning model learns the essential features to predict exact and robust salient objects. The result analysis is performed extensively and evaluated across the four benchmark RGBD datasets. The experimental results show better performance than twelve closely related state-of-the-art methods. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.