Research On RGBD Salient Object Detection Via Deep Learning

Salient Object Detection(SOD),as a fundamental task in computer vision,can help many vision tasks(such as semantic segmentation,video analysis,scene classification,etc.)to achieve better performance.Mainstream SOD methods are mostly based on RGB maps as input,while RGBD(RGB and Depth)SOD methods are methods that use both RGB maps and depth maps as input,and RGBD SOD methods have achieved better performance in challenging scenes and been rapidly developed in recent years.There are still some problems in RGBD SOD,where both RGB and depth data may have poor imaging quality,resulting in the existing methods not achieving a good performance.In addition,most of the existing RGBD SOD models are based on dualstream architecture,which neglects computational cost,making it difficult to achieve a balance between performance and computational convenience.To address the above problems within RGBD SOD,this thesis contributes as follows.(1)To cope with the problem of different quality scenes,this paper proposes an crisscross dynamic filter network including a modality-specific dynamic enhancement module(MDEM)and a scene-aware dynamic fusion module(SDFM).The MDEM generates customized dynamic filters to process RGB and depth features that can guide the model to achieve modality-specific feature enhancement.The SDFM adaptively adjusts the dynamic filter weights to select favorable cross-modal features,thus enhancing the scene focus capability of the model.We conduct extensive experiments on seven RGBD datasets to surpass the current state-of-the-art 19 RGBD SOD methods and achieve a comprehensive optimal performance,which proves the effectiveness of the proposed method.(2)To achieve a balance between computational cost and model performance in RGBD SOD,this paper proposes a depth injection framework(DIF)including an injection strategy(IS)and a depth injection module(DIM)based on a single-stream model.The IS injects the depth map into the encoder to enhance the semantic representation capability,while maintaining the computational convenience.The DIM enables low-loss alignment of depth maps and hierarchical RGB features and suppresses the interference of depth maps to ensure effective fusion.In addition,the framework has strong applicability: the framework achieves optimal performance over RGBT SOD;the DIM can improve the performance of single-stream models.Extensive experiments are conducted on six RGBD datasets,outperforming 28 state-of-the-art RGBD SOD methods,demonstrating the effectiveness of the proposed framework.