SODA²:Salient Object Detection With Structure-Adaptive & Scale-Adaptive Receptive Field
| Autor: | Min Yuan, Yan Changfei, Yuting Su, Jing Liu, Han Wang |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
General Computer Science
business.industry Computer science ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION 0211 other engineering and technologies General Engineering Pattern recognition 02 engineering and technology Salient object detection Salient objects Receptive field 0202 electrical engineering electronic engineering information engineering Dilation (morphology) 020201 artificial intelligence & image processing General Materials Science Artificial intelligence business 021101 geological & geomatics engineering |
| Zdroj: | IEEE Access. 8:204160-204172 |
| ISSN: | 2169-3536 |
| Popis: | Salient objects with complex shapes and arbitrary sizes are generally hard to detect, especially in cluttered background and complex scenes. Noticing the deficiency of single-feature-based methods, recent methods fused multiple features from Deep Neural Networks (DNNs) and obtained better performance thanks to the hierarchical feature representations. However, existing works processed the features on each semantic level with convolutions of certain fixed scale and regular sampling locations, deteriorating the predicted salient objects with complex shapes and arbitrary sizes. To tackle these problems, in this article, we propose a novel dual-adaptive salient object detection algorithm with structure-adaptive and scale-adaptive receptive field (abbreviated as SODA2). In particularly, the deformable convolution is introduced to transform the kernels’ sampling locations by augmenting extra offsets. Upon the structure-adaptive features on different semantic level, novel spatial context-aware modules (SCAMs) are devised to capture multi-scale contexts with dilated convolutions of deliberately designed dilation rates and kernel sizes. To further emphasize on saliency-relevant semantic information (e.g., humans), we proposed a region-of-interest-aware channel attention module (ROI-CAM) to redistribute the feature maps on different channels. Extensive experiments have been carried out on five challenging datasets and our algorithm performs favorably compared with state-of-the-art algorithms, validating the efficiency of the proposed models. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|