Editing Attention (section)

===Modelling===

In the domain of [[computer vision]], efforts have been made to model the mechanism of human attention, especially the bottom-up intentional mechanism<ref name="Li J, Levine MD, An X, Xu X, He H 2012">{{cite journal | vauthors = Li J, Levine MD, An X, Xu X, He H | title = Visual saliency based on scale-space analysis in the frequency domain | journal = IEEE Transactions on Pattern Analysis and Machine Intelligence | volume = 35 | issue = 4 | pages = 996–1010 | date = April 2013 | pmid = 22802112 | doi = 10.1109/TPAMI.2012.147 | arxiv = 1605.01999 | s2cid = 350786 }}</ref> and its semantic significance in classification of video contents.<ref name="Zang Wang Liu Zhang 2018 pp. 97–108">{{cite book | last1=Zang | first1=Jinliang | last2=Wang | first2=Le | last3=Liu | first3=Ziyi | last4=Zhang | first4=Qilin | last5=Hua | first5=Gang | last6=Zheng | first6=Nanning | title=Artificial Intelligence Applications and Innovations | chapter=Attention-Based Temporal Weighted Convolutional Neural Network for Action Recognition | publisher=Springer International Publishing | publication-place=Cham | volume=519 | date=2018 | isbn=978-3-319-92006-1 | doi=10.1007/978-3-319-92007-8_9 | pages=97–108}}</ref><ref name="Wang Zang Zhang Niu p=1979">{{cite journal | vauthors = Wang L, Zang J, Zhang Q, Niu Z, Hua G, Zheng N | title = Action Recognition by an Attention-Aware Temporal Weighted Convolutional Neural Network | journal = Sensors | volume = 18 | issue = 7 | pages = 1979 | date = June 2018 | pmid = 29933555 | pmc = 6069475 | doi = 10.3390/s18071979 | bibcode = 2018Senso..18.1979W | url = https://qilin-zhang.github.io/_pages/pdfs/sensors-18-01979-Action_Recognition_by_an_Attention-Aware_Temporal_Weighted_Convolutional_Neural_Network.pdf | doi-access = free }}</ref> Both [[Visual spatial attention|spatial attention]] and [[Visual temporal attention|temporal attention]] have been incorporated in such classification efforts.

Generally speaking, there are two kinds of models to mimic the bottom-up salience mechanism in static images. One is based on the spatial contrast analysis. For example, a center–surround mechanism has been used to define salience across scales, inspired by the putative neural mechanism.<ref>{{cite journal | vauthors = Itti L, Koch C, Niebur E |title=A Model of Saliency-Based Visual Attention for Rapid Scene Analysis |journal=IEEE Trans Pattern Anal Mach Intell|volume=20 |issue=11 |pages=1254–1259 |year=1998|doi=10.1109/34.730558 |citeseerx=10.1.1.53.2366 |s2cid=3108956 }}</ref> It has also been hypothesized that some visual inputs are intrinsically salient in certain background contexts and that these are actually task-independent. This model has established itself as the exemplar for salience detection and consistently used for comparison in the literature;<ref name="Li J, Levine MD, An X, Xu X, He H 2012"/> the other kind of model is based on the frequency domain analysis. This method was first proposed by Hou et al..<ref>{{Cite book |vauthors=Hou X, Zhang L |doi=10.1109/CVPR.2007.383267 |chapter-url=http://www.klab.caltech.edu/~xhou/papers/cvpr07.pdf |access-date=2015-01-10 |archive-url=https://web.archive.org/web/20150212171627/http://www.klab.caltech.edu/~xhou/papers/cvpr07.pdf |archive-date=2015-02-12 |url-status=dead |title=2007 IEEE Conference on Computer Vision and Pattern Recognition |pages=1–8 |year=2007 |isbn=978-1-4244-1179-5 |chapter=Saliency Detection: A Spectral Residual Approach |citeseerx=10.1.1.579.1650 |s2cid=15611611 }}</ref> This method was called SR. Then, the PQFT method was also introduced. Both SR and PQFT only use the phase information.<ref name="Li J, Levine MD, An X, Xu X, He H 2012"/> In 2012, the HFT method was introduced, and both the amplitude and the phase information are made use of.<ref>{{cite journal | vauthors = Li J, Levine MD, An X, Xu X, He H | title = Visual saliency based on scale-space analysis in the frequency domain | journal = IEEE Transactions on Pattern Analysis and Machine Intelligence | volume = 35 | issue = 4 | pages = 996–1010 | date = April 2013 | pmid = 22802112 | doi = 10.1109/TPAMI.2012.147 | url = http://www.cim.mcgill.ca/~lijian/06243147.pdf | arxiv = 1605.01999 | s2cid = 350786 | archive-url = https://web.archive.org/web/20130301015810/http://www.cim.mcgill.ca/~lijian/06243147.pdf | url-status = dead | archive-date = 2013-03-01 }}</ref> The Neural Abstraction Pyramid<ref>{{Cite book| vauthors = Behnke S |url=http://link.springer.com/10.1007/b11963|title=Hierarchical Neural Networks for Image Interpretation|date=2003|publisher=Springer Berlin Heidelberg|isbn=978-3-540-40722-5|series=Lecture Notes in Computer Science|volume=2766|location=Berlin, Heidelberg|doi=10.1007/b11963|s2cid=1304548}}</ref> is a hierarchical recurrent convolutional model, which incorporates bottom-up and top-down flow of information to iteratively interpret images.