Error after apply erosion filter -on images and labels/masks to- to detach two connected objects

[aymanaboghonim]

I am using pspnet and custom dataset to do binary segmentation task.
Training was running without any error but metrics of foreground object / building was lower than background.
To handle this issue, I followed the recommendations to set num_class = 1 and use_segmoid = True and reduce_label = False,
as well as using multiple loss (dice + BCE) but metrics did not improve well.
here is metrics

So, I intended to apply erosion filter on both images and labels to detach two connected objects because I think that dense object plays an important role in this situation.
After applying erosion filter using opencv kernel = np.ones((3, 3), np.uint8) cv2.erode(image,kernel,iterations = 1)
I encountered an error.

my labels is rgb.png images , here is an example

Please help , thanks in advance.