MASA-Net: Multi-level Aggregated and Self-attentive Network for Fast Salient Object Detection

With the rapid development of convolutional neural network, salient object detection has achieved satisfied visual performance. However, most previous method chooses to focus on either accuracy or speed. We propose a multi-level aggregated and self-attentive Network for fast salient object detection with a great improvement and balance of both metrics. Accordingly, we solve the scale variation problem by adopting multi-level aggregated modules which aggregate the neighboring features and enhance the feature in different resolutions. To enhance the representative feature in each layer, a self-attentive module with astrous convolution is proposed in decoder stage. In order to maintain object consistency and boundary clarity, a multi-level fusion loss which combine the first layer's boundary loss and five-layer binary cross-entropy loss is proposed. Experimental results on five widely used datasets demonstrate that our proposed method performs favorably against the state-of-the-art methods in both accuracy and speed without any pre-processing and post-processing.