Figure 10.
The image presents a convolutional neural network pipeline starting with an input of 224 by 224 by 3 that passes through a Conv 3 by 3 layer to produce 112 by 112 by 48. This is followed by M B Conv 1 3 by 3 blocks with 3 blocks at 112 by 112 by 24, then M B Conv 6 3 by 3 blocks with 5 blocks at 56 by 56 by 40. The network continues with M B Conv 6 5 by 5 blocks with 5 blocks at 28 by 28 by 64, then 7 blocks at 14 by 14 by 128, and 7 blocks at 14 by 14 by 176. A fine-tuning stage spans these deeper layers. The architecture then includes M B Conv 6 3 by 3 blocks with 3 blocks at 7 by 7 by 304, followed by an M B Conv 6 1 by 1 layer producing 7 by 7 by 512, and ends with a 7 by 7 by 2048 output.Conceptual diagram of CNN