Construction of multi-channel fusion salient object detection network based on gating mechanism and pooling network

We combine SNet network based on gating mechanism with poolnet network to solve the problem of salient object detection. The network construction of this paper is based on FPN, which is a classic U-net backbone network. Inspired by poolnet, we also introduce the global feature guidance module. By aggregating the high-level semantic information into the transposition convolution stage of different scales, the higher-level semantic features can be used more effectively. Although the introduction of global information can effectively improve the effect of saliency monitoring, how to aggregate the global and local features of different scales still needs to be further explored. Inspired by SNet network, we also integrate snet into our network. In the specific feature fusion process, the feature values of different channels are weighted, and each channel is given different weights. The more important semantic information is extracted from multiple channels, and the key semantic information in the feature map is retained. Compared with the current typical methods, we find that the introduction of snet module can reduce the generation of error areas of saliency map, and further improve the integrity of saliency map. For different regions of the same object, due to the difference of color contrast and texture, the saliency map generated by the previous method is inconsistent in the same object region. Our method can effectively solve this problem. For the same object, we can generate consistent results of saliency probability. Through quantitative evaluation with the existing 15 methods (including SOTA method). Our network can process 300 * 267 images faster than 11FPS, which is at a medium level compared with the most advanced networks. These networks include DGRL, PiCANet, PoolNet and so on. The Precision and Recall curve results show that our network performs well on DUT-O, DUT-S and ECSSD data sets, and the minimum precision values are all above 0.47. The false positive prediction of salient objects in the graph is low, and the overall performance of the model is good.