Saliency Detection via Combining Region-Level and Pixel-Level Predictions with CNNs

摘要: This paper proposes a novel saliency detection method by combining region-level estimation and pixel-level prediction with CNNs (denoted as CRPSD). For prediction, fully convolutional neural network (called CNN) is constructed modifying the VGGNet architecture to perform multi-scale feature learning, based on which an image-to-image conducted accomplish detection. estimation, adaptive superpixel region generation technique first designed partition image into regions, estimated using CNN model CNN). The saliencies are fused form final salient map another fusion And jointly learned. Extensive quantitative qualitative experiments four public benchmark datasets demonstrate that proposed greatly outperforms state-of-the-art approaches.