The model of "Image Saliency Prediction in Transformed Domain: A Deep Complex Neural Network Method", which has been published at AAAI2019.
The transformed domain fearures of images show effectiveness in distinguishing salient and non-salient regions. In this paper, we propose a novel deep complex neural network, named Sal-DCNN, to predict image saliency by learning features in both pixel and transformed domains. Before proposing Sal-DCNN, we analyze the saliency cues encoded in discrete Fourier transform (DFT) domain. Consequently, we have the following findings: 1) the phase spectrum encodes most saliency cues; 2) a certain pattern of the amplitude spectrum is important for saliency prediction; 3) the transformed domain spectrum is robust to noise and down-sampling for saliency prediction. According to these findings, we develop the structure of Sal-DCNN, including two main stages: the complex dense encoder and three-stream multi-domain decoder. Given the new Sal-DCNN structure, the saliency maps can be predicted under the supervision of ground-truth fixation maps in both pixel and transformed domains. Finally, the experimental results show that our Sal-DCNN method outperforms other 8 state-of-the-art methods for image saliency prediction on 3 databases.
If you are interested in this method please cite:
@article{jiang2019saldcnn,
title={Image Saliency Prediction in Transformed Domain: A Deep Complex Neural Network Method},
author={Lai Jiang, Zhe Wang, Mai Xu, Zulin Wang},
booktitle = {AAAI Conference on Artificial Intelligence (AAAI)},
month = {February},
year = {2019}
}
The pre-trained model can be found in dropbox. For running the demo, please downloard the model to the directory of ./model/.
This model is implemented by tensorflow-gpu 1.10.0, and the detail of our computational environment is listed in 'env.txt'. Run 'TestSALDCNN.py' to get the saliency prediction results over the images put in ./img/.
The results are output to ./result/.
Some results of our model and ground-truth.
If any question, please contact jianglai.china@buaa.edu.cn （or jianglai.china@gmail.com）, or use public issues section of this repository.
This code is distributed under MIT LICENSE.