The learning of " Background Prior-Based Salient Object Detection via Deep Reconstruction Residual" (J. Han, D. Zhang, X. Hu，and L. Guo. IEEE Transactions on Circuits and systems for video technology. Aug, 2015, pp. 1309-1321)
1 Model overview
1) Use five scales as {1/2, 1/3, 1/4, 1/5, 1/6} of the original image size to generate multiscale inputs.
2) Assume that the image boundary is mostly background. Then separately define four boundaries to get four saliency map for each image.
3) Use two-layered stacked denoising autoencoder (SDAE) model to measure the saliency.
4) Get the deep reconstruction residual maps RTOP,
5) Use the average value of the five single-scale saliency maps to generate the multiscale integrated saliency mapSR,
6) To refine saliency map, use a self-adaptive threshold t to obtain the salient cluster. Then use a Gaussian to modify the salient values according to their distances to the salient cluster center.
7) Use a graph-based segmentation algorithm to further refine saliency map.
2 Highlights
1) Use the deep reconstruction residual rather than the shallow reconstruction residual to measure the saliency. This is because the feature representation in the deep layer captures more intrinsic and latent patterns of the image boundary, which generally leads to more promising saliency detection results in line with human perception.
2) The way to refine image is unique. Instead of using the center priori, a threshold is used to obtain the salient cluster. Compare the distance between the other pixels in the whole image and the salient cluster to refine image. This method avoids the deviation caused by the salient object is not at the center.
3 Mentioned model
- Autoencoder model
 |
Feature: layer-wise pre-training
- Denoising autoencoder model (DAE)
Feature: use damaged, contaminated data to code, and can approximate the original data.
 |
- Stacked denoising autoencoder (SDAE)
 |
Feature: Superposition of multiple DAE mod