Fig. 1. Schematic diagram of VGG16-ECNN,and convolution process. (a) Schematic diagram of VGG16-ECNN; (b) convolution process
Fig. 2. Schematic diagrams of VGG16-DONN structure. (a) Simplified diagram of VGG16-ECNN; (b) 4f system [ is input complex amplitude function, is Fourier transform of , and is reconstructed signal restored by second convex lens]
Fig. 3. Weights of convolution kernels
Fig. 4. Dataset examples. (a) CelebA dataset; (b) cat and dog dataset
Fig. 5. Training flow chart
Fig. 6. Original image and output images of two network structures through partial convolution kernel in the first convolution layer.(a) Face image; (b) VGG16-ECNN; (c) VGG16-DONN
Fig. 7. Output of face image convolved with all convolution kernels in the first convolution layer
Fig. 8. Output graph of each convolution block after inputting face image to VGG16-DONN structure. (a) Conv1; (b) Conv2; (c) Conv3; (d) Conv4; (e) Conv5
Fig. 9. Original image and output graphs of two network structures through partial convolution kernel in the first convolution layer.(a) Dog image; (b) VGG16-ECNN; (c) VGG16-DONN
Fig. 10. VGG16-DONN for CDP
Fig. 11. OWM algorithm flow chart
Fig. 12. Training results of VGG16-DONN and VGG16-ECNN. (a) Training accuracy and validation accuracy of VGG16-DONN and VGG16-ECNN; (b) training loss and validation loss of VGG16-DONN and VGG16-ECNN
Fig. 13. Training/validation accuracy and time obtained by varying different parameters. (a) Weights of different layers trained by VGG16-ECNN structure; (b) weights of different layers trained by VGG16-DONN structure; (c) VGG16-DONN structure changes optimizer; (d) VGG16-DONN structure changes learning rate
Convolution layer | Time /s | Proportion of total time /% |
---|
Conv1_1 | 0.92 | 22.3 | Conv1_2 | 0.33 | 8.0 | Conv2_1 | 0.24 | 5.8 | Conv2_2 | 0.27 | 6.6 | Conv3_1 | 0.22 | 5.3 | Conv3_2 | 0.25 | 6.1 | Conv3_3 | 0.25 | 6.1 | Conv4_1 | 0.24 | 5.8 | Conv4_2 | 0.27 | 6.6 | Conv4_3 | 0.27 | 6.6 | Conv5_1 | 0.28 | 6.8 | Conv5_2 | 0.29 | 7.0 | Conv5_3 | 0.29 | 7.0 | Total time | 4.12 | 100.0 |
|
Table 1. Time of each convolutional layer of VGG16-ECNN
Network structure | Classification accuracy /% | Time /min |
---|
VGG16-DONN | 88.53 | 320 | AlexNet-DONN | 83.81 | 1117 |
|
Table 2. Classification accuracy and time of VGG16-DONN and AlexNet-DONN for cat and dog dataset
Attribute | VGG16-DONN | VGG16-ECNN | Attribute | VGG16-DONN | VGG16-ECNN |
---|
5_o_Clock_Shadow | 90.20 | 90.70 | Mouth_Slightly_Open | 62.00 | 65.30 | Arched_Eyebrows | 70.80 | 71.70 | Mustache | 96.80 | 96.30 | Attractive | 68.00 | 71.60 | Narrow_Eyes | 87.00 | 86.10 | Bags_Under_Eyes | 79.60 | 78.60 | No_Beard | 86.00 | 86.00 | Bald | 98.20 | 98.20 | Oval_Face | 71.00 | 73.40 | Bangs | 83.20 | 86.00 | Pale_Skin | 95.80 | 95.70 | Big_Lips | 67.20 | 66.20 | Pointy_Nose | 71.80 | 71.00 | Big_Nose | 81.80 | 80.30 | Receding_Hairline | 92.00 | 92.90 | Black_Hair | 75.80 | 76.00 | Rosy_Cheeks | 93.20 | 93.40 | Blond_Hair | 87.40 | 86.90 | Sideburns | 94.60 | 95.30 | Blurry | 96.20 | 94.60 | Smiling | 67.00 | 75.00 | Brown_Hair | 81.20 | 81.10 | Straight_Hair | 79.20 | 77.50 | Bushy_Eyebrows | 85.00 | 86.40 | Wavy_Hair | 66.60 | 73.60 | Chubby | 95.00 | 95.90 | Wearing_Earrings | 79.40 | 78.10 | Double_Chin | 97.40 | 95.70 | Wearing_Hat | 96.00 | 95.80 | Eyeglasses | 94.60 | 93.70 | Wearing_Lipstick | 75.20 | 81.60 | Goatee | 94.80 | 95.60 | Wearing_Necklace | 84.20 | 85.70 | Gray_Hair | 97.20 | 97.50 | Wearing_Necktie | 91.00 | 93.30 | Heavy_Makeup | 73.80 | 79.00 | Young | 77.60 | 76.90 | High_Cheekbones | 62.80 | 71.20 | Male | 77.60 | 84.90 |
|
Table 3. Accuracy of 40 kinds of face attributes obtained by VGG16-DONN and VGG16-ECNN combined with CDP module