Table 5 The architecture of the CNN
From: Performance evaluation of lung sounds classification using deep learning under variable parameters
Parameter setting | SpectroGram (SG) | Mel frequency cepstrum coefficients (MFCCs) | ||
|---|---|---|---|---|
Activations | Parameters | Activations | Parameters | |
Input | \(m\times n\times 1\) | – | \(m\times n\times 1\) | – |
2D Convolution | \(K1\times K2\times 64\) | Convolution size: 7 Stride: 1 Padding: 0 Filters number: 64 | \(J1\times J2\times 64\) | Convolution size: 3 Stride: 1 Padding: 0 Filters number: 64 |
Batch normalization | \(K1\times K2\times 64\) | – | \(J1\times J2\times 64\) | – |
ReLU | \(K1\times K2\times 64\) | – | \(J1\times J2\times 64\) | – |
Max Pooling | \(K3\times K4\times 64\) | Pooling size: 2 Stride: 1 | \(J3\times J4\times 64\) | Pooling size: 2 Stride: 1 |
Dropout | \(K3\times K4\times 64\) | 50% | \(J3\times J4\times 64\) | 50% |
Fully connected | \(1\times 1\times 10\) | Size: 10 | \(1\times 1\times 10\) | Size: 10 |
Dropout | \(1\times 1\times 10\) | 50% | \(1\times 1\times 10\) | 50% |
Fully connected | \(1\times 1\times 4\) | Size: 4 | \(1\times 1\times 4\) | Size: 4 |
Softmax | \(1\times 1\times 4\) | – | \(1\times 1\times 4\) | – |