Table 3.

Adapting (fine‐tuning) the parameters of the proposed architecture to work on tobacco images [A3 dataset (Minervini et al., 2016)] previously pre‐trained with Arabidopsis plants [A1, A2, and A4 (Bell and Dee, 2016; Minervini et al., 2016)]. We progressively increase the number of training images to find a suitable number of images required to create a meaningful model that can count tobacco leaves. The table reports the results on the held‐out testing set

No. of training images	DiCa	|DiC|a	MSEa	%b
7	−0.39 (1.65)	1.32 (1.07)	2.83	23.2
14	0.00 (1.32)	0.96 (0.90)	1.75	32.1
21	0.27 (1.36)	0.87 (1.07)	1.91	41.1
27	0.25 (1.20)	0.86 (0.87)	1.50	37.5

DiC, difference in count; |DiC|, absolute DiC; MSE, mean squared error; %. percentage difference.

^aBest values are those closer to 0.

^bBest values are those closer to 1 (i.e. 100%).