Hello,
Yes, we can have N+1 filters also to include background class.

Actually, we can treat the background class in two ways:

1. The classification head is a single label classifier(uses softmax), hence we need to have N+1 filters, N=number of classes + 1 for background.
2. The classification head is a multi-label classifier(uses sigmoid), hence we can have only N filters. When the network outputs all 0s, we treat it as background.