Table 1 A summary of main results of the image processing and network compression.
From: Deep learning-enabled mobile application for efficient and robust herb image recognition
DNN strategy | Top1 Mean ± std (%) | Top5 Mean ± std (%) |
|---|---|---|
a. Image processing | ||
Small DNN w/o image processing | 64.65 ± 1.09 | 86.02 ± 0.51 |
Small DNN + image processing | 68.89 ± 0.69 | 88.03 ± 0.69 |
b. Network pre-train | ||
Small DNN w/o pre-train | 46.51 ± 0.90 | 75.38 ± 1.48 |
Small DNN + pre-train | 68.89 ± 0.69 | 88.03 ± 0.69 |
c. Network transfer | ||
Large DNN | 70.61 ± 1.00 | 88.97 ± 0.41 |
Small DNN | 68.89 ± 0.69 | 88.03 ± 0.69 |
Small DNN + transfer | 70.97 ± 0.50 | 88.72 ± 0.79 |
Small DNN w/o pre-train + transfer | 55.55 ± 3.22 | 79.25 ± 2.11 |
d. Network cut | ||
Small DNN + transfer | 70.97 ± 0.50 | 88.72 ± 0.79 |
Small DNN + cut(\(\alpha =0.5\), \(\lambda =0\)) | 67.62 ± 0.84 | 85.69 ± 0.85 |
Small DNN + cut (\(\alpha =0.5\), \(\lambda =1\)) | 69.44 ± 1.10 | 87.66 ± 0.73 |
Small DNN + transfer + cut (\(\alpha =0.5\), \(\lambda =0\)) | 68.01 ± 0.47 | 85.88 ± 0.54 |
Small DNN + transfer + cut (\(\alpha =0.5\), \(\lambda =1\)) | 69.98 ± 1.09 | 87.71 ± 0.73 |
e. Network structure | ||
Small DNN + transfer + cut (\(\alpha =0.5\), \(\lambda =1\)) | 69.98 ± 1.09 | 87.71 ± 0.73 |
New small DNN | 62.66 ± 1.19 | 84.91 ± 1.06 |
New small DNN + transfer (\(\lambda =1\)) | 66.68 ± 1.21 | 86.44 ± 0.78 |
