Free Access

Table 3.

Prediction accuracy on the validation and test sets is listed for each model.

Accuracy
Training
Network type Model Weights Validation Test Duration Steps Time (s)
Capsule 1 9 861 010 80% 78% 14h 7860
2 4 945 042 84% 86% 15h 7692 0.35
3 4 724 992 50% 50% 2h 11990 0.013
4 3 428 222 88% 88% 11h 2965 0.79
5 3 320 722 82% 82% 11h 3250 0.83
ALED-m 216 608 88% 92% 4h 1620 0.63
6 117 280 88% 92% 2h 1950 0.22
7 13 880 88% 86% < 1h 3960 0.022
8 5984 88% 90% < 1h 9200 0.015
9 2816 84% 90% < 1h 2320 0.0085
10 2546 86% 84% < 1h 2015 0.0086
Convolutional 11 22 848 778 Over-fitting < 1h 456
12 4 551 906 88% 86% < 1h 1653 0.0029
13 4 551 906 82% 86% < 1h 360 0.0029
14 4 551 906 60% 66% < 1h 234 0.0029
15 875 586 66% 60% < 1h 1000 0.0016

Notes. The time taken to train each model is given under the ‘Duration’ column. The total number of times the weights were updated during training is given under the ‘Steps’ column. The time taken to classify a single image is given under the ‘Time’ column, and this number was averaged over 50 images.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.