CCL-Net: Complete Comprehensive Learning and Modality Preserving-Based RGBD Complex Salient Object Detection

Abstract

Recently, a significant advancement in salient object detection has been achieved by utilizing depth modality and cross-complementary features fusion in a convolution neural network. The existing models usually compute RGB and depth features separately and fuse them to find saliency. The recent complementary-based features fusion model only focused on cross-complementary features and ignored modality-dependent non-complementary features. The complementary and non-complementary features are essential to predict the correct salient object. The proposed model targets the above limitations and achieves them by proposing three-stream networks. The color and depth streams are used to produce modalities-dependent non-complementary features. The fused stream utilized the salient features from both streams to explore complementary features. The global deep guidance loss function is formulated to purify the deep localized saliencies features used to start the fusion process. The output of this fused stream combines with the other two streams to preserve the non-complementary features to generate final saliency. The experimental analysis using recent metrics across the four publicly available challenging RGBD datasets is thoroughly evaluated. The performance is compared with thirteen RGBD state-of-the-art saliencies, which shows better performance in comparison with the closely related state-of-the-art methods. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.