Open Access

Table 2

Performance of interpolation methods and of the proposed ANNs, with and without the removal of outlier from the training set.

Method Error factor Memory Speed

mean 99th per. max (MB) ms)
No outlier removal near. neighbor ×13.1 ×11.3 ×3e5 1650 62
linear 15.7 ×2.3 ×143 1650 1.5e3

spline linear 15.7 ×2.3 ×144 1650
cubic 11.2 ×2.2 ×122 1650
quintic 19.1 ×2.9 ×304 1650

RBF linear 10.2 96.8 ×99 1650 1.1e4
cubic 10.4 ×2.1 ×112 1650 1.1e4
quintic 10.9 ×2.1 ×118 1650 1.1e4

ANN R 7.3 64.8 ×81 118 12
R+P 6.2 49.7 ×84 118 13

Outlier removal on training set near. neighbor ×13.1 ×11.6 ×3e5 1650 62
linear 15.9 ×2.4 ×143 1650 1.5e3

spline linear 15.9 ×2.4 ×144 1650
cubic 11.1 ×2.2 ×120 1650
quintic 20.0 ×2.7 ×285 1650

RBF linear 10.3 97.3 ×97.5 1650 1.1e4
cubic 10.5 ×2.0 ×106 1650 1.1e4
quintic 10.9 ×2.0 ×114 1650 1.1e4

ANN R 5.1 42.0 ×32.8 118 12
R+P 5.5 42.3 ×41 118 13

R+P+C 4.9 44.5 ×44 51 14
R+P+D 4.5 33.1 ×33.8 125 11
R+P+C+D 4.8 37.9 ×37.6 43 14

Notes. Evaluation speeds are measured on the full set of L lines for 1000 random points. The measurements are performed on a personal laptop equipped with eight logical cores running at 3.00 GHz. Error factors are evaluated on the test set. For neural network architectures, C stands for a line clustering and specialist networks, D for a dense architecture, P for a polynomial transform and R for the design of the last hidden layer using PCA. For each criterion, the best obtained values are highlighted in bold.

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.