7. Depthwise Separable Convolution layer
K
K
F
F
N M
F
F
F:
K:
N,M:
N
!"#×%"×& !"#×&
Convolution layer
K
K
1
N
F
F
1
1
N
MN
'(×)(×* + *×)(×, !"# + #&
K=3 , N=3 , F=256 , M=128
Depthwise Separable
3"×3×256"×128 = 226,492,416 3"×256"×3 + 3×256"×128 = 26,935,296
3"×128×3 = 3,456 3"×3 + 3×128 = 411
Depthwise Conv
14. •
VGG16:16 InceptionResNetv2:572
Deep Pneumonia Net:575
• InceptionResNetv2
Depthwise
Separable Convolution
0
20
40
60
80
100
1
74.2
93.8 98.1
Accuracy(%)
Classifier performance on the test data
VGG16 InceptionResNetv2 Deep Pneumonia Net
15. Confusion matrix for Deep Pneumonia Net Confusion matrix for Inception Resnet v2
Accuracy:96.5% Accuracy:92.0%
Recall:97.3% Recall:94.4%
223 11
11 379
206
368
28
22
truelabel
truelabel
predicted labelpredicted label