Page 179 - Kaleidoscope Academic Conference Proceedings 2020

Industry-driven digital transformation

Table 1 – Network parameters

Layer (type) Output shape Param #
conv2d_18 (Conv2D) (None,32,32,64) 640
max_pooling2d_17 (MaxPooling) (None,16,16,64) 0
batch_normalization_14 (Batc) (None,16,16,64) 256
flatten_17 (Flatten) (None,16384) 0
dense_33 (Dense) (None,256) 4194560
dense_34 (Dense) (None,4) 1028
Total params: 4196484
Trainable params: 4196356
Non-trainable params: 128

Table 2 – Object recognition results when epoch=5 (GAN training, gun: 118, knife: 572, other: 139, scissor: 43)
Evaluation results
Gun Knife Other Scissor Total Recall Average recall
Gun 53 1 4 0 58 0.913793
Knife 3 269 12 1 285 0.94386
Input 0.6828397
Other 9 3 57 0 69 0.826087
Scissor 0 19 1 1 21 0.047619
Total 65 292 74 2 433
Precision 0.815385 0.921233 0.77027 0.5 Accuracy rate
Average Precision 0.751722 0.8775982

affected by the change of epoch number. Moreover, the 433 and 172 images for GAN training at CASE 1 and CASE 2
recall when epoch number=5 is shown in Table 2. From respectively). Moreover, the recall when epoch number=5 is
this table, the knife obtains the highest recall (0.94386) since shown in Table 3. From this table, recall of scissors increases
GAN training uses the highest proportion of knife (66%). to 0.47619 (the recall of scissors is 0.047619 at CASE 1)
Although gun and other do not have a very high proportion since the proportion of scissors for GAN training increased
for GAN training (13% and 16% respectively), it still obtains from 5% to 25%. Also, the proportion of suspicious objects
a good recall, 0.913793 and 0.826087 respectively. Only for GAN training is almost the same, it’s difficult to obtain
scissor has a low recall of 0.047619 and we speculate that it extremely low accuracy.
maybe caused by the low proportion of scissor during GAN
training (5%). 4.3 CASE 3

4.2 CASE 2 Objective: CASE 1 and CASE 2 confirmed the proportion
of images employed for GAN training affects the accuracy rate
of CNN. However, both of these two cases used the images
Objective: This part will evaluate whether the proportion
generated by GAN for CNN training. In this part, we will
of images used for GAN training affect the accuracy rate of
evaluate whether the GAN affects the accuracy rate of CNN.
CNN.
Method: In this part, we will use the same methods as for
Method: As with CASE 1, there are still four categories
CASE 1 and CASE 2. The difference is that we use not only
of real experimental images that will be employed for GAN
the GAN images but also original real experimental images
training, including gun, knife, other, and scissor. However,
for CNN training to verify whether the GAN affects the CNN
we fix 43 images/categories for GAN training. Here we still
accuracy rate. Regarding CNN training without using GAN
choose the epoch as 1000 and batch_size is 32. After that, the
images at CASE 1, we will use the real experimental images
GAN generates 1000 guns, 1000 knives, 1000 other objects,
(118 guns (13%), 572 knives (66%), 139 other objects (16%),
and 1000 scissors for CNN training. To evaluate the accuracy
43 scissors (5%); total 872 images) by active/passive imagers
rate, we use the same number of real experimental images
directly for CNN training. Meanwhile, the same images with
with CASE 1 (58 guns, 285 knives, 69 other objects, and 21
CASE 1 (58 guns (13%), 285 knives (66%), 69 other objects
scissors) for CNN evaluation.
(16%), and 21 scissors (5%)) will be used for CNN evaluation.
In addition, CASE 2 is the same way.
Result: The red line in Figure 6 shows the accuracy rate of
CNN evaluation when we use 43 guns, 43 knives, 43 other
objects, and 43 scissors for GAN training. Compared with
CASE 1, the overall accuracy decreases from approximately Result: Figure 7 and Figure 8 illustrate the accuracy rate
87% to 68% when we use less images for GAN training (total when GAN images and original experimental images are used

– 121 –