Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
SEOUL | Oct.7, 2016
RhD / CEO 강현인, Oct.7, 2016
IMPLEMENTATION OF DEEP LEARNING ALPR IN TK1
10.57 cm
2
AGENDA
1. Jetson TK1 module
2. Expanded Peripherals
3. ALPR requirement I/O
4. ALPR Procedure
5. Deep Learning for ALPR
6. Image Samples in Real Environment
7. Performance in Deep Learning
8. GUI Application
3 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
1. JETSON TK1 MODULE NVIDIA JETSON TK1
4 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
2. EXPANDED PERIPHERALS EMBEDDED BOARD
5 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
3. ALPR REQUIREMENT I/O (1) CLCU (CAMERA LENS CONTROL UNIT)
6 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
3. ALPR REQUIREMENT I/O (2)
주차관제 시스템 구성
7 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
4. ALPR PROCEDURE (1)
ALPR Conventional Approach
IMAGE ACQUISITION
RoI based iris control
PREPROCESSING
Contrast Enhancement
8 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
4. ALPR PROCEDURE (2)
ALPR Conventional Approach
LP DETECTION
Find coarse position of LP
LP NORMALIZATION
Find fine-position of LP
Apply affine transform to make LP into rectangle
9 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
4. ALPR PROCEDURE (3)
ALPR Conventional Approach
LP TYPE RECOGNITION
Classify the type of LP
CHARACTER SEGMENTATION
Find coarse position of character
CHARACTER RECOGNITION
Recognize each character
8,4,마,7,4,1,4
10 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
4. ALPR PROCEDURE (4)
LP DETECTION
Find the position of LP
LP RECOGNITION
Recognize whole characters at once
84마7414
IMAGE ACQUISITION
RoI based iris control
ALPR End-to-End Approach of Deep Learning
11 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
5. DEEP LEARNING FOR ALPR (1)
“CNN in object detection & recognition
Deep convolutional neural network already exceeds human-level accuracy of
detection. And, still the advance in CNN is ongoing.”
FASTER R-CNN(2015) YOLO(2016) SINGLE SHOT MULTIBOX DETECTOR(SSD, 2016)
What makes the difference
with Convolutional Neural Network and Deep Learning
12 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
5. DEEP LEARNING FOR ALPR (2)
C++ based library: good for integrating with existing system.
Caffe Model Zoo: You can get lots of pre-trained models for free including some of latest paper’s works which are not officially published yet.
And, Nvidia’s DIGITS supports Caffe!
You can train/test a Caffe model with just few clicks and a little bit of typing using DIGITS’ GUI
Deep learning framework: Caffe / DIGITS
13 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
5. DEEP LEARNING FOR ALPR (3)
1. Get lots of training data.
2. Design a model for your task.
3. Train the model and evaluate your model with DIGITS.
4. Go back to 3 until you are satisfied with your test result.
5. Write a code for applying the trained model on your product using Caffe’s library (C++ or Python).
6. Watch how it works in real world!
Deep learning development process with Caffe & DIGITS
14 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
6. IMAGE SAMPLES IN REAL ENVIRONMENT (1) <인식 영상>
15 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
6. IMAGE SAMPLES IN REAL ENVIRONMENT (2) <미인식 영상>
16 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
7. PERFORMANCE IN DEEP LEARNING
17 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
8. GUI APPLICATION (1)
RECOGNITION
18 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
8. GUI APPLICATION (2)
RECOGNITION
19 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
8. GUI APPLICATION (3)
SPC CONFIG
20 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
8. GUI APPLICATION (4)
CAMERA
21 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
8. GUI APPLICATION (5)
LOG
22 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
8. GUI APPLICATION (6)
CENTER
SEOUL | Oct.7, 2016
THANK YOU