BINARY step
Many components of the Audiveris engine are meant to work on a black & white input image. In such image, the black pixels compose the foreground and the white pixels the background.
The goal of the BINARY step is to provide this black & white image.
Inputs
- The gray image of the sheet
- The chosen binarization filter – by default, this is the adaptive filter
Output
- The black & white image of the sheet
Processing
The gray value – a number between 0 and 255 – of each image pixel is compared with a threshold value. The result is:
- black (0) if gray value <= threshold value
- white (255) if gray value > threshold value
What is this threshold value? It depends on the binarization filter used on the sheet image.
Via the Book → Set book parameters pull-down menu, the precise filter and its parameters can be defined for the default, book and sheet scopes.
The GLOBAL filter
This is the older filter.
It provides a single threshold value which is applied globally on every pixel of the gray image.
The threshold default value is 140, and we can change this value via the Book → Set book parameters pull-down menu.
It generally provides good results, except when the input image exhibits different illumination values. Let’s consider the BachInvention5.png image, available in the data/examples folder of Audiveris.
Here is the bottom part of the gray image:
Note that the image appears dark on the left and pale on the right.
If we apply the global filter on it, we get this binary image:
The following engine steps will miss both the left part and the right part of this system. The left part because almost every pixel looks like foreground. The right part because several staff lines lack foreground pixels.
In this case, increasing or decreasing the overall threshold value has no point, because it cannot improve both parts at the same time.
We need something more sensitive to the local environment.
The ADAPTIVE filter
This is the newer filter.
Instead of being a constant value, a specific threshold value is now computed for every pixel, based on its location in the image.
The actual gray value is measured for every pixel in a rectangular neighborhood around the current location. Based on this population, the engine computes two values:1
mean: the mean gray value,std_dev: the standard deviation of gray values.
The resulting threshold is then defined according to the following formula:
threshold = mean_coeff * mean + std_dev_coeff * std_dev
The parameters are by default 0.7 for mean_coeff and 0.9 for std_dev_coeff. We can modify them via the Book → Set book parameters pull-down menu.
On the same example, we now get this binary image: 
Our experience on hundreds of gray images has shown that the adaptive filter gives good results and its default parameters are OK. Therefore, the adaptive filter has been promoted as the default binarization filter for every input…
… until this discussion was recently posted on Audiveris forum.
Here below is a portion of the gray input image posted in the discussion:
We can notice that the staff lines, as well as the bar lines, are painted in a pale gray color.
Here, the pale gray is not a transition between black and white as seen on the other scores.
Applying the standard adaptive filter gives this binary result:
Here, we must go back to the global filter.
With a threshold value set to 225, we now obtain this binary result:
Boards
Binarization board
The Binarization board is present in the Binary tab.
It allows the end user to interactively choose the filter, adjust its parameters, and observe the results on the image immediately.
Please refer to the Binarization board section.
Binary board
Available in the Gray and Binary tabs, the Binary pixel board presents the resulting black or white pixel value at the current location.
Please refer to the Binary board section.
-
The implementation does not use brute force to compute mean and standard deviation around each and every point, but the result is the same. ↩