一种面向移动端的图像风格迁移模型压缩算法 
下载： 1039次
An Image Style Transformation Model Compression Algorithm for Mobile Terminal
图 & 表
图 1. 模型迭代优化风格迁移算法整体结构
Fig. 1. Overall structure of model iterative optimization style transformation algorithm
下载图片 查看原文
图 2. 图像风格迁移网络结构
Fig. 2. Image style transformation network structure
下载图片 查看原文
图 3. 改进前、后的网络结构。(a)改进前;(b)改进后
Fig. 3. Network architecture before and after improvement. (a) Before improvement; (b) after improvement
下载图片 查看原文
图 4. 网络模型压缩前、后风格迁移评分。(a)压缩前;(b)压缩后
Fig. 4. Style transformation scores before and after network model compression. (a) Before compression; (b) after compression
下载图片 查看原文
图 5. 网络模型压缩后不同风格的图像迁移评分
Fig. 5. Transformation scores of image with different styles after network model compression
下载图片 查看原文
表 1各层网络的参数量
Table1. Number of network parameters at each layer
| Layer | Number ofparameters | Ratio of numberof parameters /% |
|---|
| Conv1 | 7776 | 0.4 | | Conv2 | 18432 | 1.1 | | Conv3 | 73728 | 4.4 | | Res1 | 294912 | 17.6 | | Res2 | 294912 | 17.6 | | Res3 | 294912 | 17.6 | | Res4 | 294912 | 17.6 | | Res5 | 294912 | 17.6 | | Deconv1 | 73728 | 4.4 | | Deconv2 | 18432 | 1.1 | | Deconv3 | 7776 | 0.4 |
|
查看原文
表 2模型压缩前、后参数量的对比
Table2. Comparison of number of parameters before and after model compression
| Layer | Number of parameters | Ratio ofcompression /% |
|---|
| Beforecompression | Aftercompression |
|---|
| Conv1 | 7776 | 7776 | 0 | | Conv2 | 18432 | 18432 | 0 | | Conv3 | 73728 | 73728 | 0 | | Res1 | 294912 | 53248 | 81.9 | | Res2 | 294912 | 53248 | 81.9 | | Res3 | 294912 | 53248 | 81.9 | | Res4 | 294912 | 53248 | 81.9 | | Res5 | 294912 | 53248 | 81.9 | | Deconv1 | 73728 | 73728 | 0 | | Deconv2 | 18432 | 18432 | 0 | | Deconv3 | 7776 | 7776 | 0 | | Total | 1674432 | 363712 | 78.3 |
|
查看原文
表 3压缩前、后模型相关数据
Table3. Data related to the compressed model
| Model | Before compression | After compression | Compressionratio | Speed-upratio |
|---|
| Size /MB | | | Time /ms | Size /MB | Time /ms |
|---|
| Wave | 21.1 | 6937 | 4.4 | 4590 | 4.56 | 1.51 | | Feathers | 21.1 | 6851 | 4.4 | 5314 | 4.56 | 1.29 | | Mosaic | 21.1 | 6950 | 4.4 | 4421 | 4.56 | 1.57 | | Denoised | 21.1 | 4420 | 4.4 | 3571 | 4.56 | 1.23 |
|
查看原文
表 4移动端测试结果
Table4. Test results at mobile terminal
| Brand | Model | Installationtime /s | Initializationtime /s | CPUusage /% |
|---|
| Huawei | P10 Plus | 3.85 | 0.83 | 1.00 | | Huawei | Honor 9 | 3.31 | 0.72 | 0.13 | | Huawei | Honor 7X | 4.38 | 1.16 | 0.82 | | 360 | N5S | 5.71 | 0.97 | 1.07 | | Vivo | V173CA | 5.15 | 1.56 | 1.78 | | Sony | Xperia XA | 9.47 | 1.38 | 4.27 |
|
查看原文
裴斐, 刘进锋, 李崤河. 一种面向移动端的图像风格迁移模型压缩算法[J]. 激光与光电子学进展, 2020, 57(6): 061021. Fei Pei, Jinfeng Liu, Xiaohe Li. An Image Style Transformation Model Compression Algorithm for Mobile Terminal[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061021.
PDF全文
