Training, export and test Customized YOLOv5
Dataset NUMBERS Computer Vision
The result can be seen: Android App

The dataset used in this work did not contain small numbers usually denoting parts of the whole.
To isolate them from the total mass, you can use the calculation of the area of the predicted bonding boxes.
To do this, in the new_sorter function of the model, you need to add something like that:

x1 = non_max[:, 0]
y1 = non_max[:, 1]
x2 = non_max[:, 2]
y2 = non_max[:, 3] 

areas = (x2 - x1) * (y2 - y1)

Further, you can work with the corresponding values both in the function itself and on the device.