Skip to main content

Table 3 DenseNet-169 for tongue image quality control

From: Tongue image quality assessment based on a deep convolutional neural network

Layers

Feature map size

Structure

Convolution

200 × 200

7 × 7 conv, 32, stride 2

Pooling

100 × 100

3 × 3 max pool, stride 2

Dense Block (1)

100 × 100

\(\left[ {\begin{array}{*{20}l} {1 \times 1} \hfill & {{\text{conv}},} \hfill & {128} \hfill \\ {3 \times 3} \hfill & {{\text{conv}},} \hfill & {32} \hfill \\ \end{array} } \right] \times 6\)

Transition Layer (1)

100 × 100

1 × 1 conv, 112

50 × 50

2 × 2 average pool, stride 2

Dense Block (2)

50 × 50

\(\left[ {\begin{array}{*{20}l} {1 \times 1} \hfill & {{\text{conv}},} \hfill & {128} \hfill \\ {3 \times 3} \hfill & {{\text{conv}},} \hfill & {32} \hfill \\ \end{array} } \right] \times 12\)

Transition Layer (2)

50 × 50

1 × 1 conv, 248

25 × 25

2 × 2 average pool, stride 2

Dense Block (3)

25 × 25

\(\left[ {\begin{array}{*{20}l} {1 \times 1} \hfill & {{\text{conv}},} \hfill & {128} \hfill \\ {3 \times 3} \hfill & {{\text{conv}},} \hfill & {32} \hfill \\ \end{array} } \right] \times {\text{32}}\)

Transition Layer (3)

25 × 25

1 × 1 conv, 636

12 × 12

2 × 2 average pool, stride 2

Dense Block (4)

12 × 12

\(\left[ {\begin{array}{*{20}l} {1 \times 1} \hfill & {{\text{conv}},} \hfill & {128} \hfill \\ {3 \times 3} \hfill & {{\text{conv}},} \hfill & {32} \hfill \\ \end{array} } \right] \times {\text{32}}\)

Classification Layer

1 × 1

12 × 12 global average pool

2208D fully connected layer with ReLU

2D fully connected layer

Softmax

  1. In DenseNet-169, the growth rate is set to 32, each bottleneck layer produces 128 feature maps, and the reduction is 0.5