Top-down Feature Aggregation Block Fusion Network for Salient Object Detection

The emergence of deep neural networks and full convolutional neural networks has brought great progress to salient object detection. In this paper, we propose a new type of deep full convolutional neural network structure, named top-down feature aggregation block fusion network, which aims to fuse the rich features of feature aggregation blocks at each layer. In addition to the features of this layer, the feature aggregation blocks have other layer features, that is, each layer of feature aggregation blocks has both strong semantic information of the deep network and detailed features of the shallow network. In the top-down fusion process, the residual information of each layer can be learned like ResNet. At the same time, a non-local attention mechanism is introduced to improve the relevance of the context, and multiple auxiliary supervision connections are added to the intermediate layers, so that the network can more easily optimize and accelerate convergence. We have performed experiments on six benchmark datasets, and the results of the experiments show that our model is superior to the state-of-the-art methods both quantitatively and qualitatively.