Table 3.

Composition of the DICCN model.

Type	Patch size/stride	Output size
Deep separable convolution	3 × 3/1	224 × 224 × 64
Max-pooling	3 × 3/2	112 × 112 × 64
Deep separable convolution	3 × 3/1	112 × 112 × 64
Max-pooling	3 × 3/2	56 × 56 × 64
Batch normalization	–	56 × 56 × 64
Inception	–	56 × 56 × 576
Max-pooling	3 × 3/2	28 × 28 × 576
Inception	–	28 × 28 × 576
Inception	–	28 × 28 × 576
Inception	–	28 × 28 × 576
Inception	–	28 × 28 × 576
Max-pooling	3 × 3/2	14 × 14 × 576
Inception	–	14 × 14 × 960
Max-pooling	3 × 3/2	8 × 8 × 960
GAP	–	960
Softmax	–	7