International Core Journal of Engineering 2020-26 | Page 86

澳
model using 320 real images. The relevant experimental
results are shown in Table II.
for model training. (3) SynImg-32Cls+RealImg: Pre-train the
model using 3 200 synthetic images, then fine-tune the
T ABLE II R ESULTS OF THE SYNTHESIS ALGORITHM AND COMPARISON WITH [12].
Method
Setting: Adidas Aldi Apple Becks BMW Carls Chim Training/Test Corona DHL Erdi Esso Fedex Ferra Ford Fost
Image Split Google Guin Hein HP Milka Nvid Paul Pepsi
(per logo class)
RealImg([12])
SynImg-32Cls([12])
SynImg-32Cls
+ RealImg᷉[12]᷊
RealImg
(This paper)
SynImg-32Cls᷉This
paper᷊
Coke
Ritt Shell Sing Starb Stel Texa Tsin Ups
23.7 57.5 63.0 69.6 63.7 50.6 55.2 26.8
Training: 10 RealImg 79.0 25.8 61.2 44.2 45.9 80.6 64.3 43.2
Test: 60 RealImg 47.7 58.2 61.8 21.3 19.4 17.4 48.2 17.8
34.8 45.8 71.8 70.2 79.6 56.7 56.9 52.2
9.4 47.3 9.6 70.3 39.9 28.3 15.8 21.7
Training: 100 SynImg 6.1 11.1 4.1 44.7 22.9 60.9 43.6 28.8
Test: 60 RealImg 23.0 16.7 43.1 9.9 4.6 1.1 39.1 9.7
22.7 38.3 15.5 65.6 28.7 55.1 27.4 20.1
26.8 63.7 65.8 72.7 81.3 52.7 63.6 30.0
Training: 100 SynImg 76.0 31.5 63.0 52.2 54.3 90.0 84.0 46.6
+10 RealImg 58.0 52.6 65.2 23.2 24.0 12.5 54.1 23.6
Test: 60 RealImg 37.9 45.6 75.0 73.8 79.0 64.2 57.4 54.4
24.4 57.2 66.6 72.0 70.8 42.8 55.3 24.8
Training: 10 RealImg 82.8 29.5 62.5 44.1 42.7 87.2 59.3 39.9
Test: 60 RealImg 51.2 54.6 65.1 24.2 15.5 16.9 52.3 17.4
32.3 44.7 72.7 69.7 77.3 62.5 52.6 44.3
38.6 54.0 34.6 44.1 17.6 18.4 33.9 20.6
Training: 100 SynImg 8.2 21.5 12.1 49.6 16.6 28.2 35.4 50.3
Test: 60 RealImg 46.4 40.0 33.1 26.1 9.9 18.8 64.5 22.3
39.4 45.5 15.8 55.5 17.6 47.5 38.7 39.9
51.1 62.1 29.8
31.0 69.0 74.3 78.0 76.6 Training: 100 SynImg 88.5 35.3 70.9 53.5 49.0 87.9 76.9 52.9
+10 RealImg 61.7 66.9 69.3 34.1 25.8 26.2 67.8 22.8
Test: 60 RealImg 41.8 45.5 79.4 72.5 79.7 69.2 62.1 61.5
36.2 61.3 30.1 46.9 25.3 30.8 32.2 22.3
Training: 100 SynImg 11.4 16.7 19.6 47.1 28.0 33.2 33.8 50.1
RealImg᷉fusion᷊ +10 RealImg 45.7 41.6 37.1 25.6 13.0 16.3 63.7 20.9
᷉This paper᷊ Test: 60 RealImg 33.4 45.0 14.1 62.7 22.8 46.4 43.5 36.6
Training: 100 SynImg 31.1 71.2 71.6 71.7 84.4 48.0 66.3 28.5
RealImg᷉fusion᷊ +10 RealImg 84.4 33.3 82.0 51.7 51.0 90.3 76.4 54.8
+RealImg
᷉This paper᷊ Fine-tune˖10 RealImg 59.6 67.8 69.4 37.8 23.2 22.9 66.0 23.6
Test: 60 RealImg 40.6 46.0 75.3 77.2 83.5 67.8 63.6 64.2
SynImg-32Cls
+ RealImg
(This paper)
SynImg-32Cls +
SynImg-32Cls +
mAP
50.4
27.6
54.8
50.5
32.6
58.5
34.2
58.9
which fully verifies the effectiveness of the synthesis
algorithm. It is worth mentioning that the improvement in
performance is not dependent on additional manual labeling,
nor does it require the construction of a large (463 categories)
common set of logo exemplars as in [12].
In Table II, simply using 320 real images for training,
RealImg (This paper) obtained similar experimental results
to RealImg ([12]) (50.5% VS. 50.4%), although 50.5% is a
result of one experiment, actually for RealImg this paper has
carried out several experiments, and the mAP is changed
very little at around 50.0%. The reason why the experimental
results are unstable is mainly due to the randomness of the
neural network algorithm itself (such as the optimization
algorithm of Faster R-CNN using random gradient descent).
Therefore, this paper basically reproduces the experimental
results of [12] on FlickrLogos-32. On this basis,
SynImg-32Cls(This
paper)
and
SynImg-32Cls+RealImg(This paper) using the proposed
synthesis algorithm have obvious advantages compared with
the method of [12] (32.6% VS. 27.6%, 58.5% VS. 54.8%),
In addition, this paper has also observed that although
SynImg-32Cls(This paper) is greatly improved compared to
SynImg-32Cls ([12]), the effect of using only synthetic
images for training is still far from the method of using a
small number of real images. The potential cause of this
situation may be that there is a large distribution difference
between the real image and the synthetic image, and the
details learned by the model on the synthetic image are
difficult to generalize into the real image. From this point of
view, the key to the synthetic data extension training set
64