Get best of the week

HoughNet: Search for vanishing points fused with a classic algorithm


While dozens and even hundreds of proven architectures of artificial neural networks (ANNs) are trained in the world of object recognition, warming the planet with powerful video cards and creating a “panacea” for all tasks of computer vision, we are firmly following the research path in Smart Engines, offering new effective ANN architectures to solve specific problems. Today we will talk about HafNet - a new way to search for vanishing points on images.


Hough transformation and its quick implementation


. , , , , . ( ). () : , , , .. , , , , , .


(xi,yi). , yi=xia+b a b. b=-xia+yi ab, , (xi,yi) . : , , , . : , . ( – , ).


, , , – .


, . .



, , : – O(n3), n – .


() – , , () . O(n2 log(n)), . , , , [5]. , : : « » ( H1), « » ( H2), « » ( H3) « » ( H4). , H12 H34 , .



( , ). , , . .


. , , - , : ( , , ), – . , . . , - . , H12 , . , , H34 , . , , H12 , , , . ( , H12 ).



( ). , ( – ).


, , : ? , … , !


HoughNet


, , - (, – , – ). «» ( , ). «» . «» , , ?


, ( HoughNet), , - . , , – , , . , ( [1]).


1conv12 filters 5x5, stride 1x1, no paddingrelu
2conv12 filters 5x5, stride 2x2, no paddingrelu
3conv12 filters 5x5, stride 1x1, no paddingrelu
4FHTH12 for vertical, H34 for horizontal-
5conv12 filters 3x9, stride 1x1, no paddingrelu
6conv12 filters 3x5, stride 1x1, no paddingrelu
7conv12 filters 3x9, stride 1x1, no paddingrelu
8conv12 filters 3x5, stride 1x1, no paddingrelu
9FHTH34 for both branchesg-
10conv16 filters 5x5, stride 3x3, no paddingrelu
11conv16 filters 5x5, stride 3x3, no paddingrelu
12conv1 filter 5x5, stride 3x3, no padding1 – rbf

: «» , «» . .


. MIDV-500 [6]. , . 50 . ( , 30 ) , – .


, , ICDAR 2011 dewarping contest dataset ( 100 - , ) .



«» ( ), state-of-the-art [7] [8].


[7][8]HoughNet
31.3%49.6%50.1%59.7%


[1] Sheshkus A. et al. HoughNet: neural network architecture for vanishing points detection // 2019 International Conference on Document Analysis and Recognition (ICDAR). – 2020. doi: 10.1109/ICDAR.2019.00140.
[2] . ., . ., . . // . – 2014. – . 64. – №. 3. – . 25-34.
[3] .. : . … . . .-. . – ., 2019. – 24 .
[4] [ ]: . . – : https://ru.wikipedia.org/wiki/_/ ( : 13.03.2020).
[5] Nikolaev D. P., Karpenko S. M., Nikolaev I. P., Nikolayev P. P. Hough Transform: Underestimated Tool in the Computer Vision Field // 22st European Conference on Modelling and Simulation, ECMS 2008. – Nicosia, Cyprys, 2008. – P. 238–243.
[6] Arlazarov V. V. et al. MIDV-500: a dataset for identity document analysis and recognition on mobile devices in video stream // . – 2019. – . 43. – №. 5.
[7] Y. Takezawa, M. Hasegawa, and S. Tabbone, “Cameracaptured document image perspective distortion correction using vanishing point detection based on radon transform,” in Pattern Recognition (ICPR), 2016 23rd International Conference on. IEEE, 2016, pp. 3968–3974.
[8] Y. Takezawa, M. Hasegawa, and S. Tabbone, “Robust perspective rectification of camera-captured document images,” in Document Analysis and Recognition (ICDAR), 2017 14th IAPR International Conference on, vol. 6. IEEE, 2017, pp. 27–32.

More articles:


All Articles