International Core Journal of Engineering 2020-26 | Page 85

澳
designed with their main color, and in terms of shape and
semantic relevant random background image, thus realizing
outline, logo is generally not presented as a regular
the synthesis of the logo image. Unlike the random
rectangular structure. (As shown in Fig. 4). Therefore, at the
placement of the logo exemplar in [12, 16], this paper pays
data level, this paper adopts the Main Color Similarity
special attention to where the logo exemplar appears in the
Comparison Algorithm. Algorithm 1 gives the main
background. Specifically, except for a few logos (such as
algorithm flow of the logo image synthesis step.
Google) composed of plain text symbols, most logos are
Algorithm 1: Logo image synthesis based on main color similarity comparison
Input: Transformed logo exemplar I * , Randomly selected background image B ;
Output: Synthesized logo image B * ;
1) Randomly select ROI in B for plating I * with Shape ( I *) Shape ( ROI ) ;
2) for every pixel p in I * and
ROI
judgeHSVColor ( p ) ;
end for
3) Count max 2 Color ( I *) and
max 2 Color ( ROI ) ;
overlap
(
max
2
Color
(
I
*),
max
2 Color ( ROI ))
4) if
Randomly select another background image B ;
Return 1).
else
Return B * B . replacedWith ( ROI , I *) .
end if
Combining the above four processes, the synthesis
algorithm can not only realize the automatic synthesis of
large-scale logo images based on context, but also ensure the
accurate annotation of each logo object in the synthetic
image without any omission. Examples of synthetic logo
images of this paper are shown in Fig. 7.
Fig. 7. Examples of the synthetic image based on the synthesis algorithm.
This paper choose mAP for performance evaluation of the
algorithm.
III. E XPERIMENTS AND R ESULTS
A. Datasets
The dataset this paper used in the experiment is
FlickrLogos-32. FlickrLogos-32 was designed for logo
retrieval and multi-class logo detection and object
recognition. There are 32 logo classes with 70 images each
class. During the actual training process, this paper strictly
follows the official classification criteria, and divided only
10 images of each logo class as training samples, and the
remaining 60 samples were used as test samples.
C. Experiments and Analysis
First, this paper basically reproduces the experimental
results of [12] on the FlickrLogos-32. Specifically, this paper
uses the Faster R-CNN as the algorithm framework for logo
recognition. The network structure used is VGG-16, and also
its pre-trained model on the PASCAL VOC 2007.
Parameters and training strategies remain the same as [12].
Based on the recurring results, this paper uses the synthesis
algorithm to automatically generate 100 synthetic images
together with corresponding annotations for each logo class,
and then train logo recognition model based on the synthetic
images and the real images.
B. Evaluation Metrics
Object recognition usually require certain evaluation
metrics to evaluate the performance of the algorithm. The
commonly used evaluation metrics in the field of logo
recognition is mAP (mean Average Precision). The mAP
comprehensively characterizes the precision and recall. The
larger the value, the better the performance of the algorithm.
In general, the comparative experiments with [12] mainly
have the following three processes. (1) RealImg: Only 320
labeled real images are used for model training. (2)
SynImg-32Cls: Only 3200 labeled synthetic images are used
63