Table 2. Effects of changing the overall architecture on performance and runtime.

Type Index	Architectural Change	Training Time/Speedup	Processing Rate/Speedup	Validation Accuracy
1	Overall architecture	13.9 h/1.0×	3.88 ms/1.0×	0.932
2	Removed layer FC7	12.8 h/1.1×	3.40 ms/1.1×	0.908
3	Removed layers FC6, 7	12.5 h/1.1×	3.05 ms/1.3×	0.897
4	Removed layers FC6, 7, Conv5	10.1 h/1.4×	2.45 ms/1.6×	0.869
5	Removed layers FC6, 7, Conv4, 5	9.5 h/1.5×	1.78 ms/2.2×	0.724
6	Removed layers FC6, 7, Conv3, 4, 5	6.6 h/2.1×	1.66 ms/2.3×	0.633
7	Removed layers FC6, 7, Conv2, 3, 4, 5	3.1 h/4.5×	1.65 ms/2.3×	0.630
8	Adjust Layers Conv2, 3, 4, 5: 192, 320, 320, 192 filters	12.0 h/1.2×	3.48 ms/1.1×	0.929
9	Adjust Layers Conv2, 3, 4, 5: 128, 256, 256, 128 filters	9.2 h/1.5×	2.84 ms/1.4×	0.924
10	Adjust Layers Conv2, 3, 4, 5: 64, 192, 192, 64 filters	7.1 h/2.0×	2.28 ms/1.7×	0.917

The processing rate indicates the time of a feed-forward pass for one image. Speedup denotes the time ratio of the changed architectures versus the overall architecture.