Bi-attention network for bi-directional salient object detection

Saliency detection models based on neural networks have achieved outstanding results, but there are still problems such as low accuracy of object boundaries and redundant parameters. To alleviate these problems, we make full use of position and contour information from the down-sampling layers, and optimize the detection result layer by layer. First, this paper designs an attention-based adaptive fusion module (AAF), which can suppress the background and highlight the foreground that is more relevant to the detection task. It automatically learns the fusion weights of different features to filter out conflict information. Second, this paper proposes a bi-attention block module which combines reverse attention and positive attention. Third, this paper introduces bi-directional task learning by decomposing the image into high-frequency and low-frequency components. This approach fully exploits the complementary and independent nature of different frequency information. Finally, the proposed method is compared with 14 state-of-the-art methods on 6 datasets, and achieves very competitive results. Additionally, the model size is only 114.19MB, and the inference speed can reach nearly 40 FPS.