Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Report on ALDPR

Automatic License Plate Detection and


Recognition: A Comparison of Architectures
2 INTRODUCTION
1 ABSTRACT This is more of a three-part
Automatic License Plate introduction. We won’t go deeply
Detection and Recognition into working of the models, but
systems have become one of the rather delve into it enough so that
most progressive areas in the field we can understand their
of Intelligent Traffic Management comparisons
Systems (ITML). So as a result,
there have been quite a lot of
detection and recognition pipelines These are the models we
that have been touched upon took into consideration for the
mildly. This slapdash development detection pipeline: -
in the technology has led to some 1. YOLOv4
sloppy outputs, thus making the 2. VGG16
same a bit unreliable. Further
These two models employ a
scrutiny revealed, the detection
CNN based approach to their
pipeline is the most vulnerable
algorithm, which prompted the
when efficiency is taken into the
need of a non-CNN model to have
question. So, in this paper, we
a fair comparison. This is where the
undertake the burden of finding
inbuilt toolboxes of MATLAB lent
the best real-time detection
us a huge hand.
algorithm coupled with a state of
the art and easy to implement For the recognition pipeline
Recognition pipeline of the CNN models, we just used
the easyOCR library in python,
which does a perfect job.

TEAM10 1
Report on ALDPR

3 PROPOSED employ a CNN based algorithm,


and that is what YOLO does too, to
METHODOLOGIES
put it lightly. To understand the
3.1 YOLOV4 working of YOLO, we first need to
It is Common Knowledge any understand the working of the
image detection techniques will single state object detectors

Figure 1

As you can see in Figure 1, In trained using the respective python


single stage, we apply the codes for training and testing
detection head directly on the
Then it is just a simple
feature map
matter of running the detector
Now, to access the features, with local images for detection.
we need a cloud-based IDE,
preferably Google Collab. We will
then clone in a darknet server to
remotely access the detector. We
would recommend mounting the
google drive to facilitate faster
runtime, but using the local
machine will work too, albeit a bit
less smoothly.

Image datasets can then be


zipped and uploaded and can be

TEAM10 2
Report on ALDPR

4 MATLAB AND IMAGE characters and saves them as a


PROCESSING new template in MATLAB memory,

Unlike the traditional way of -Letter Detection which reads the


using neural networks for ALDPR, characters from given image and
this method uses image processing finds the most matching
toolbox from MATLAB to process alphanumeric characters and
the image through multiple
-Plate Detection which processes
number of algorithms to provide an
the image and calls the above two
alpha numeric conversion of the
programs to detect the characters
image into a text format.
in the license plate.
The concept we use here for
Choosing the image we want
detecting number plates consists
and running the MATLAB code, It
of three programs,
will give us the plate number
-Template Creation which within a matter of few seconds
converts images of alphanumeric with fair accuracy

TEAM10 3
Report on ALDPR

5 VGG16

5.1 STRUCTURAL CONFIGURATION

TEAM10 4
Report on ALDPR

5.2 TABULAR COLUMN FOR THE CONFIG

TEAM10 5
Report on ALDPR

5.3 SEQUENCES OF STEPS FOR • A VGG 16 based network has


DETECTION USING VGG-16 been trained for 80% of the
data and with 50 epochs, the
• A data set of 433 images and accuracy varies from 0.36 to
their pixel size and 0.94 at the final stages
segmented coordinates for • Then the remaining 0.2 data
the bounding box has been are tested, the final
established. bounding boxes with images
• An example of bounding box are displayed from the
coordinates has been trained vgg-16 network.
accessed and an array of
data are created

TEAM10 6
Report on ALDPR

6 COMPARISON OF ARCHITECTURES

PIPELINE MODEL TIME FOR EFFICIENCY


OUTPUT

YOLOv4 1.4 sec (without GPU) 91%


0.72 sec (with GPU)

DETECTION

VGG 16 Runs on internet-based 86.21%


notebook, so to train it
takes around 50 mins,
to test around 7ms per
step so 7 seconds for
20% of 433 images

easyOCR 1.6 sec (without GPU) 82%


0.9 sec (with GPU)
RECOGNITION

TEAM10 7
Report on ALDPR

ARCHITECTURE TIME FOR EFFICIENCY


OUTPUT

YOLOv4 with easyOCR (single 1.7 sec (without 91%


pipeline) GPU)

1.1 sec (with GPU)

MATLAB 3.8 sec 72%

errors in detection of the


7 INFERENCE
concerned images. The range of
We can clearly see that the the license plate in the image is
VGG16 model requires retraining of very important factor and this
images, so it becomes obsolete to impacts the model quite a bit
go through with it, even though it
The YOLOv4 trained
has the fastest output.
easyOCR model clearly tested out
As the MATLAB model is not the charts in terms of detection
based on a CNN algorithm, but and output speed, which gave it an
rather a position based edge over the other models
architecture, it is prone to basic

8 CONCLUSION
Thus we used a YOLOv4
model with easyOCR for the
ALDPR

TEAM10 8

You might also like