Open Access

Street View House Number Identification Based on Deep Learning


Cite

Figure 1.

Network structure
Network structure

Figure 2.

Optimizer effect
Optimizer effect

Figure 3.

Training Accuracy of different regularization
Training Accuracy of different regularization

Figure 4.

Training Loss of different regularization
Training Loss of different regularization

Figure 5.

Test Accuracy of different regularization
Test Accuracy of different regularization

Figure 6.

Tetst Loss of different regularization
Tetst Loss of different regularization

Figure 7.

SVHN-Complete house number
SVHN-Complete house number

Figure 8.

SVHN-Part number
SVHN-Part number

Figure 9.

Example of train set
Example of train set

Figure 10.

Example of extra set
Example of extra set

Figure 11.

Example of test set
Example of test set

Figure 12.

Category distribution of SVHN
Category distribution of SVHN

Figure 13.

Training of the model after adding data augmentation
Training of the model after adding data augmentation

Figure 14.

Figure 1 Test result
Figure 1 Test result

OPTIMIZER PARAMETER SETTING

Optimizerparameter
SGDlr=0.001,
Adamlr=0.001,
Adamaxlr=0.002,
RMSproplr=0.001,

AUGMENTATION RESULT

Subset categoryNumber of samples
Training set73257
Extra set531131
Test set26032

OPTIMIZERS TRAINING RESULTS

optimizerTop Accuracy/%
SGD87.350184
Adam89.090000
Adamax88.955900
RMSprop88.676000

RESULT AFTER DATA AUGMENTATION

Train sample numbertest sample numberBest test accuracytime
732572603290.012291h24min
6043882603292.324836h17min

RESULT OF DIFFERENT WEIGHT DECAY

weight_decayTrain Acc%(e)Test Acc%(e)
0.0189.01538(85)87.14659(85)
0.00591.59397(57)88.95590(57)
0.002594.51470(88)90.01229(88)
0.00197.21119(90)89.70498(24)
eISSN:
2470-8038
Language:
English
Publication timeframe:
4 times per year
Journal Subjects:
Computer Sciences, other