Advanced Hough-based on-device document localization
Advanced Hough-Based On-Device Document Localization. In an embodiment, lines are detected in an input image of a document. The lines are searched for candidate quadrilaterals. For at least a subset of the found candidate quadrilaterals, a contour score is calculated, and the candidate quadrilaterals are saved or discarded based on their contour scores. For each saved candidate quadrilateral, a contrast score is calculated. A final candidate quadrilateral is selected, based on the combined contour and contrast scores for the saved candidate quadrilaterals, to represent the borders of the document.
1 . A method comprising using at least one hardware processor to:
receive an input image of a document, the input image comprising pixel data captured by an image sensor;
detect a plurality of edges in the document by processing the pixel data of the input image;
detect a plurality of lines based on the plurality of edges using a computer-implemented line detection algorithm;
search the plurality of lines for combinations of the plurality of lines that represent a plurality of new candidate quadrilaterals;
for each of at least a subset of the plurality of new candidate quadrilaterals,
calculate a contour score for each new candidate quadrilateral,
when a saved set of candidate quadrilaterals has not yet reached a predetermined maximum size K, add said new candidate quadrilateral to the saved set, and,
when the saved set of candidate quadrilaterals has reached the predetermined maximum size K,
when the contour score of said new candidate quadrilateral is not greater than a contour score of any candidate quadrilateral in the saved set, discard the new candidate quadrilateral, and,
when the contour score of said new candidate quadrilateral is greater than a contour score of at least one candidate quadrilateral in the saved set, discard a candidate quadrilateral with a lowest contour score from the saved set, and add said new candidate quadrilateral to the saved set, such that the saved set never exceeds the predetermined maximum size K;
after the search, for each candidate quadrilateral in the saved set,
calculate a contrast score for each candidate quadrilateral, and
calculate a combined score from the contour score and the contrast score calculated for each candidate quadrilateral; and
select a candidate quadrilateral having a highest combined score from the saved set to represent borders of the document.
2 . The method of claim 1 , wherein the input image is a lower resolution version of an original image, and wherein the method further comprises using the at least one hardware processor to refine the selected candidate quadrilateral by:
extracting a region that represents each border in the selected candidate quadrilateral;
applying a similarity transformation to the selected candidate quadrilateral to produce a transformed quadrilateral, wherein the similarity transformation
makes each of the extracted regions, representing a horizontal border, horizontal,
makes each of the extracted regions, representing a vertical border, vertical, and
scales the selected candidate quadrilateral to a higher resolution; and
detecting lines in the transformed quadrilateral.
3 . The method of claim 1 , wherein the search comprises:
searching for combinations of four lines from the plurality of lines; and
searching for combinations of three lines from the plurality of lines, and, for each combination of three lines that is found, interpolating a fourth line based on the three lines and a known aspect ratio of the document.
4 . The method of claim 1 , further comprising using the at least one hardware processor to, for each of the plurality of new candidate quadrilaterals, before calculating the contour score for the new candidate quadrilateral:
determine whether or not an inverse image of the new candidate quadrilateral differs from an ideal rectangle representing the document by a predefined threshold difference; and,
when the inverse image of the new candidate quadrilateral differs from the ideal rectangle by the predefined threshold difference, discard the new candidate quadrilateral without calculating the contour score for the new candidate quadrilateral and without adding the new candidate quadrilateral to the saved set.
5 . The method of claim 4 , wherein determining whether or not the inverse image of the new candidate quadrilateral differs from the ideal rectangle by the predefined threshold difference comprises calculating whether an aspect ratio of the inverse image of the new candidate quadrilateral differs from an aspect ratio of the ideal rectangle by more than a predetermined percentage.
6 . The method of claim 4 , wherein determining whether or not the inverse image of the new candidate quadrilateral differs from the ideal rectangle by the predefined threshold difference comprises calculating whether a difference between an angle of the inverse image of the new candidate quadrilateral and an angle of 90 degrees exceeds a predetermined number of degrees.
7 . The method of claim 1 , wherein the contour score is calculated as:
Contour
Score
=
∑
{
b
}
w
(
b
)
1
+
∑
{
b
}
(
1
-
c
(
b
)
)
-
∑
{
b
}
w
′
(
b
)
wherein {b} are the combination of lines in the new candidate quadrilateral,
wherein w(b) is a total edge intensity inside a segment b,
wherein w′(b) is a total edge intensity of segments that lie on a straight line containing segment b, do not intersect each other, have one common point within segment b, and have at least the threshold length, and
wherein c(b) is a fraction of non-zero edge pixels inside segment b.
8 . The method of claim 7 , further comprising using the at least one hardware processor to, for each of the plurality of new candidate quadrilaterals, before calculating the contour score for the new candidate quadrilateral:
determine whether or not each of w(b), w′(b), and c(b) has a value that is outside a respective range; and,
when determining that any of w(b), w′(b), and c(b) have a value that is outside the respective range, discard the new candidate quadrilateral without calculating the contour score for the new candidate quadrilateral and without adding the new candidate quadrilateral to the saved set.
9 . The method of claim 1 , wherein, for each candidate quadrilateral in the saved set, the contrast score is calculated based on an χ 2 distance between red-green-blue (RGB) pixel histograms of inner areas and outer areas of the candidate quadrilateral.
10 . The method of claim 9 , wherein, for each candidate quadrilateral in the saved set, the contrast score is calculated after projectively normalizing the candidate quadrilateral.
11 . The method of claim 1 , wherein detecting the plurality of edges comprises generating a horizontal edge map for primarily horizontal edges and a vertical edge map for primarily vertical edges.
12 . The method of claim 11 , wherein generating each of the horizontal edge map and the vertical edge map comprises:
applying morphological filtering to the input image;
calculating derivative values along a respective axis of the input image;
averaging the derivative values;
suppressing non-maxima having absolute derivative values greater than one;
collecting connectivity components; and
filtering the collected connectivity components, based on size, to produce remaining edges,
wherein the respective edge map comprises the remaining edges.
13 . The method of claim 12 , wherein generating each of the horizontal edge map and the vertical edge map further comprises blurring the remaining edges using a Gaussian function.
14 . The method of claim 11 , wherein detecting the plurality of lines comprises:
detecting primarily horizontal lines from the horizontal edge map; and
detecting primarily vertical lines from the vertical edge map.
15 . The method of claim 14 , wherein detecting the plurality of lines comprises, for each of the horizontal edge map and the vertical edge map:
applying a Fast Hough Transform to one or more regions of the respective edge map;
determining a global maximum across all of the one or more regions based on the Fast Hough Transform;
selecting up to a predetermined number of local maxima within each of the one or more regions based on the global maximum; and
applying an inverse of the Fast Hough Transform to the one or more regions to convert the local maxima into straight lines.
16 . The method of claim 15 , wherein selecting up toa predetermined number of local maxima within each of the one or more regions based on the global maximum comprises selecting a local maximum only if that local maximum exceeds a threshold percentage of the global maximum and that local maximum lies more than a threshold number of pixels away, by an l 2 norm, from all previously selected local maxima.
17 . The method of claim 1 , wherein the combined score is a linear combination of the contour score and the contrast score.
18 . The method of claim 1 , wherein the at least one hardware processor is comprised in a mobile device.
19 . A system comprising:
at least one hardware processor; and
software configured to, when executed by the at least one hardware processor,
receive an input image of a document, the input image comprising pixel data captured by an image sensor;
detect a plurality of edges in the document by processing the pixel data of the input image;
detect a plurality of lines based on the plurality of edges using a computer-implemented line detection algorithm;
search the plurality of lines for combinations of the plurality of lines that represent a plurality of new candidate quadrilaterals,
for each of at least a subset of the plurality of new candidate quadrilaterals,
calculate a contour score for each new candidate quadrilateral,
when a saved set of candidate quadrilaterals has not yet reached a predetermined maximum size K, add said new candidate quadrilateral to the saved set, and,
when the saved set of candidate quadrilaterals has reached the predetermined maximum size K,
when the contour score of said new candidate quadrilateral is not greater than a contour score of any candidate quadrilateral in the saved set, discard the new candidate quadrilateral, and,
when the contour score of said new candidate quadrilateral is greater than a contour score of at least one candidate quadrilateral in the saved set, discard a candidate quadrilateral with a lowest contour score from the saved set, and add said new candidate quadrilateral to the saved set, such that the saved set never exceeds the predetermined maximum size K;
after the search, for each candidate quadrilateral in the saved set,
calculate a contrast score for each candidate quadrilateral, and
calculate a combined score from the contour score and the contrast score calculated for each candidate quadrilateral; and
selecta candidate quadrilateral having a highest combined score from the saved set to represent borders of the document.
20 . A non-transitory computer-readable medium having instructions stored therein, wherein the instructions, when executed by a processor, cause the processor to:
receive an input image of a document, the input image comprising pixel data captured by an image sensor;
detect a plurality of edges in the document by processing the pixel data of the input image;
detect a plurality of lines based on the plurality of edges using a computer-implemented line detection algorithm;
search the plurality of lines for combinations of the plurality of lines that represent a plurality of new candidate quadrilaterals;
for each of at least a subset of the plurality of new candidate quadrilaterals,
calculate a contour score for each new candidate quadrilateral,
when a saved set of candidate quadrilaterals has not yet reached a predetermined maximum size K, add said new candidate quadrilateral to the saved set, and,
when the saved set of candidate quadrilaterals has reached the predetermined maximum size K,
when the contour score of said new candidate quadrilateral is not greater than a contour score of any candidate quadrilateral in the saved set, discard the new candidate quadrilateral, and,
when the contour score of said new candidate quadrilateral is greater than a contour score of at least one candidate quadrilateral in the saved set, discard a candidate quadrilateral with a lowest contour score from the saved set, and add said new candidate quadrilateral to the saved set, such that the saved set never exceeds the predetermined maximum size K;
after the search, for each candidate quadrilateral in the saved set,
calculate a contrast score for each candidate quadrilateral, and
calculate a combined score from the contour score and the contrast score calculated for each candidate quadrilateral; and
select a candidate quadrilateral having a highest combined score from the saved set to represent borders of the document.