Salient Object Detection via Feature Permutation and Space Activation

Salient object detection occupies an important position in the field of computer vision. How to deal with feature information on different scales becomes the key to obtain excellent prediction results. Two contributions are made in this article. On the one hand, a feature permutation method for salient object detection is proposed. The proposed method is a convolutional neural network based on the self-encoding network structure. It uses the concept of scale representation proposed in this paper to group and permute the multiscale feature maps of different layers in the neural network. So the proposed method obtains a more generalized salient object detection model and a more accurate prediction results about salient object detection. On the other hand, the proposed method adopts the double-conv residual and FReLU activation for the output of the model, so that more complete pixel information could be obtained, and the spatial information is also activated as well. The characteristics of the two algorithms are fused to act on the learning and training of the model. Finally, the proposed algorithm is compared with the mainstream salient object detection algorithms, and the experimental results show that the proposed algorithm obtains the best results from all.