Research Article

HCRCaaS: A Handwritten Character Recognition Container as a Service Based on QoS Guarantee Algorithm

Table 1

Structure, recognition accuracy, and hyperparameter settings of three proposed DCNN-based models and other comparison models: each proposed DCNN-based model contains 1 input layer, multiconvolutional layers, multimax pooling layers, and 1 fully connected layer.

IndexModel structureAccuracy (%)Hyper parameters

196 × 96-80C3 × 3-MP2 × 2-160C2 × 2-MP2 × 2-240C2 × 2-MP2 × 2-320C2 × 2-MP2 × 2-400C2 × 2-MP2 × 2-480C1 × 1-512FC-375595.12minBatch = 128, iterNum = 300,000 dropout (fc) = 0.5,
: reducing 0.1 every 70,000 iterations
296 × 96-100C3 × 3-MP2 × 2-200C2 × 2-MP2 × 2-300C2 × 2-MP2 × 2-400C2 × 2-MP2 × 2-500C2 × 2-600C1 × 1-512FC-375595.80minBatch = 128, iterNum = 300,000 dropout (fc) = 0.5,
: reducing 0.1 every 70,000 iterations
396 × 96-96C3 × 3-MP3 × 3-128C3 × 3-MP3 × 3-160C3 × 3-MP3 × 3-256C3 × 3-256C3 × 3-MP3 × 3-384C3 × 3-384C3 × 3-MP3 × 3-1024FC-375597.30minBatch = 128, iterNum = 300,000 dropout (fc) = 0.5,
: reducing 0.1 every 70,000 iterations
4CNN-Fujitsu [34]94.77
5ART-CNN [34]95.04
6HCCR-Gradient-GoogLeNet [38]96.28
7HCCR-Ensemble-GoogLeNet [38]96.64
8Multi-CNN voting [39]96.79
9R-CNN-voting [40]95.55
10ATR-CNN voting [40]96.06
11MQDF-THU [51]92.56
12MQDF-HIT [52]92.61
13DLQDF [2]92.72

Settings for input layer (“Input”) are given in rows, which present size of input picture (“size × size”). Settings for each convolutional layer (“Conv”) are given in rows, which show number of output feature maps and receptive field size (“numCsize × size”). Settings of max pooling (MP) are specified by type and kernel size (“Mpsize × size”). Settings of fully connected (FC) are specified by dimensionality (“FCdimensionality”).