Gated Contextual Features for Salient Object Detection

The effective extraction of local and contextual visual cues carrying information of different scales is crucial for accurate detection of the salient object(s) with varying shape, size, and location. The atrous spatial pyramid pooling (ASPP) and its dense versions are widely used for extracting contextual features for dense prediction tasks. The skip connections in densely or moderately connected ASPP directly propagate the context information from a parallel dilated convolution to the next higher rate dilated convolution to combat the “gridding issue” in atrous convolutions. The aggregated context from several scales may dilute features belonging to small objects or confuse between the salient object and the background. To emphasize invariance features for different scale visual patterns in an image, a gate-based context extraction module is proposed in this work. Gate functions are embedded in the interbranch short connection of the proposed module. The learnable gates are deployed to decide on the relevance of the contextual information extracted at a lower scale for the next higher scale. Experimental results on salient object detection tasks demonstrate that gates are helpful to retain relevant contextual information across multiple-scales of the context-extraction module. The performance of the proposed gated contextual feature-based salient object detector is evaluated on five broadly used saliency detection benchmarks by comparing it with the other 13 state-of-the-art approaches. Experimental outcomes show that the proposed method achieves a favorable performance for various compared evaluation measures.