A multi-level feature weight fusion model for salient object detection

Although the Fully Convolutional Neural Networks (FCNs) has achieved good performance in salient object detection, there are problems, such as fuzzy boundary and unsatisfactory performance in complex scenes. Hence, how to better integrate multi-level convolution feature requires further investigation. This paper proposes a salient object detection algorithm, which uses Gram matrix and its F norm to weigh the importance of each multi-level feature map and uses weight to fuse multi-level prediction results recursively, finally generate the final saliency map. The algorithm evaluates the importance of different depth multi-level feature maps by calculating the Gram matrix's F norm of feature tensor slices. The multi-level feature maps are fused effectively according to the weight. It reduces the loss of multi-level prediction results during fusion, and preserves the spatial details. Besides, to achieve a more accurate boundary, a deep supervision is used to optimize salient feature maps’ results. Pixel-level supervision information from ground truth will guide each layer’s prediction. Experiments on five benchmark data sets demonstrate that the proposed method performs well in various scenes, especially in complex scenes.