Extract time from image of a clock
Extraction of time from a given image of a clock using Canny edge detection algorithm and Hough transformation.
Function extract_time extracts the minutes and hours.
Function extract_time_bonus extracts seconds if the needle exists, if not, outputs -1.
IMPLEMENTATION
Extracting hours and minutes:
Input images are converted from RGB to grayscale, because knowledge of the color is redundant. Than the images are filtered with Gaussian filter in order to remove some artifacts and noise and to make gradients of the seconds needle smaller than the gradients of hours needle.
The main problem is how to determine of all of the edges, extracted by Canny edge detection algorithm (you can check out my Matlab implemenation from scratch here), which edges belong to needles. To make this possible we will only observe the middle part of the images. That is the part where we possibly only have needles edges.
So the Canny edge detection algorithm is applied only on the cropped part that contains 25% of the dimensions around the middle pixel of the images:
Then the Hough transformation is applied on the extracted edges. The transformation is tuned so that it extracts 2 longest lines:
Now we have positions of the 2 needles, but we can not determine which one is for hours and which one is for minutes, because we maybe have cropped both needles and now they have the same length. The idea is to extract edges on the original image and to apply Hough transformation to get some number (hyperparameter) of longest lines (edges). Then from the longest lines find 2 lines that match the lines extracted from the cropped part. The longer line will be minutes needle and the shorter one will be hours needle.
Using basic math knowledge the angles of the lines are transformed into the angles that are measured from the vertical line that passes thorugh the 6h and 12h in clockwise direction. Resulting angles are converted into minutes and hours.
Extracting seconds:
The idea is simillar as before. Canny edge detection is applied on the cropped part of the image, but now it is applied 2 times. First time the threshold is set high so that the seconds needle is not extracted. Second time the threshold is set low so that all of the needles are extracted. First image is then subtracted from the second image. The resulting image contains only the seconds needle. This is ilustrated on clock7.png:
Then the Hough transoformation is applied and tuned to extract only single longest line:
The angle of the line is converted into the seconds same as for the hours and minutes.
TEST
Run the main.m to test the functions on 12 images.
NOTE
The hyperparameters in Hough's transformation are tuned for Matlab 2017.