Advanced topic in computer engineering

<span class="text_page_counter">Trang 1</span><div class="page_container" data-page="1">

HO CHI MINH CITY UNIVERSITY OF TECHNOLOGY AND EDUCATION FACULTY FOR HIGH QUALITY TRAINING

FINAL REPORT

Advanced Topic in Computer Engineering

Ho Chi Minh City, June 2023

</div><span class="text_page_counter">Trang 2</span><div class="page_container" data-page="2">

1.4. STRUCTURE OF THE REPORT ... 2

CHAPTER 2 IDENTIFY OBJECTS WITH REAL-TIME OBJECT DETECTORS . 3 2.1. INTRODUCTION OF OBJECT RECOGNITION TECHNOLOGY ... 3

2.2.3.1. Extended efficient layer aggregation networks ... 7

2.2.3.2. Model scaling for concatenation-based models ... 8

2.2.4. Trainable bag-of-freebies ... 9

2.2.4.1. Planned re-parameterized convolution ... 9

2.2.4.2. Coarse for auxiliary and fine for lead losss ... 10

CHAPTER 3 DATABASE MANAGEMENT APPLICATIONS USING C# ... 12

</div><span class="text_page_counter">Trang 3</span><div class="page_container" data-page="3">

3.1. INTRODUCTION TO DATABASE AND STRUCTURED QUERY

3.2 BUILD DATABASE MANAGEMENT APPLICATIONS USING C# SOFTWARE, WPF PROGRAMMING AND MICROSOFT SQL ...15

CHAPTER 4 RESEARCHING TO AUTOMOTIVE EMBEDDED WITH HELLA VIETNAM ... 22

4.1. INTRODUCTION TO HELLA COMPANY ...22

4.2. SOFTWARE ENGINEER IN FORVIA HELLA ... 22

4.2.1. Software Engineer in FORVIA Hella ...22

4.2.2. How do we test in Hella? ... 22

4.2.3. Responsibilities of Tester Engineer ... 23

4.2.4. To become a Tester Engineer in Hella ... 23

4.2.5. Software Quality Assurance ... 23

4.3. CAREER OPPORTUNITIES ... 23

4.4. SOME PHOTOS OF THE WORKSHOP ... 24

CHAPTER 5 CONCLUSION ... 25

</div><span class="text_page_counter">Trang 4</span><div class="page_container" data-page="4">

TABLE OF FIGURES

Figure 2.1: Comparison with other real-time object detectors ... 5

Figure 2.2: Compound scaling with EfficientNet ... 7

Figure 2.3: Extended efficient layer aggregation networks ...8

Figure 2.4: Model scaling for concatenation-based models ... 8

Figure 2.5: Planned re-parameterized model ... 9

Figure 2.6: Coarse for auxiliary and fine for lead head label assigner. ...10

Figure 3.1: Management program block of Mobile ...16

Figure 3.2: The interface “ĐANG NHAP” ...16

Figure 3.3: The interface “QUAN LY MOBILE” ... 17

Figure 3.4: The interface “SAN PHAM” ...18

Figure 3.5: The interface “LOAI SAN PHAM” ... 19

Figure 3,6: The interface “NHA CUNG CAP” ... 20

Figure 3.7: The interface “KHACH HANG” ... 21

Figure 4.1: Workshop Automotive Trends ... 24

</div><span class="text_page_counter">Trang 5</span><div class="page_container" data-page="5">

ABBREVIATION

AI Artificial Intelligence

SQL Structured Query Language NPU Neural Processing Units WPF Windows Presentation Foundation DBMS Database Management Systems CNN Convolutional Neural Network

</div><span class="text_page_counter">Trang 6</span><div class="page_container" data-page="6">

CHAPTER 1 INTRODUCTION

1.1. INTRODUCTION

1.1.1. General

This subject introduces topics on different fields such as AI, Databases, Embedded, and Microchips, ... for students to have an overview and review the knowledge they have learned.

1.1.2. Objective

This course allows students to refresh necessary specialist information while also understanding the real-world working environment. Additionally, students get access to engineering-related job possibilities through firms.

1.2. CONTENT TOPIC 1:

Name of the topic: Identify objects with real-time object detectors Presentation: Le Thi Kieu Giang and Nguyen Hung Thinh

Content: This topic aims to provide students with the foundational knowledge of artificial intelligence and applications, providing learners with the basics of pattern recognition and machine learning.

The report provides some information about YOLOv7. TOPIC 2:

Name of the topic: Database management applications using C# Presentation: Le Thi Kieu Giang and Nguyen Hung Thinh

Content: This topic implements database management systems. From a user perspective, the course will discuss conceptual data modeling, physical data modeling, data computation, schema design, database querying, and database manipulation.

TOPIC 3:

Name of the topic: Researching to automotive embedded with Hella Vietnam Presentation: Le Thi Kieu Giang and Nguyen Hung Thinh

Content: This topic is a workshop from Hella Vietnam company. The

</div><span class="text_page_counter">Trang 7</span><div class="page_container" data-page="7">

design, implementation including analysis of embedded system hardware and software. Design, implement, and debug complex software applications on embedded systems. Real-time operating system base for real-time control embedded systems.

1.3. SKILLS AND KNOWLEDGE TO BE ACHIEVED AFTER COMPLETING THE SUBJECT

After completing the course, students can:

- The course also equips skills in programming artificial intelligence applications, using Python language, and building identification applications.

- Program a management application and build a database using C#, WPF. - Knowledge on the design of embedded systems including the design, implementation including analysis of embedded system hardware and software. 1.4. STRUCTURE OF THE REPORT

Content 1: Identify objects with real-time object detectors Content 2: Database management applications using C#

Content 3: Researching to automotive embedded with Hella Vietnam

</div><span class="text_page_counter">Trang 8</span><div class="page_container" data-page="8">

Machine vision is one of the key facets of artificial intelligence. The field of computer vision encompasses techniques for digital image acquisition, processing, analysis, and recognition, object detection, image generation, and image hyper resolution, among other things. Because it is applied so frequently in daily life, object detection is probably the most profound aspect of machine vision.

The ability of computer systems and software to find and identify objects in an image is known as object detection. Object detection has been extensively used in security systems, autonomous vehicles, pedestrian counting, face recognition, and vehicle detection. Object recognition has numerous applications in a wide range of fields of study and practice. Like any other technology, a variety of cutting-edge 2.1.2. Development trends

Nowadays, business operations are closely tied to IT, and businesses are competing to invest in IT. The field of information technology has long used the term artificial intelligence (AI). In addition, AI describes the application of algorithms to the completion of specific tasks through the analysis of vast amounts of data to derive statistical generalizations or estimates. Through the use of these algorithms, a computer program can learn, reason, and make decisions in a manner similar to that of the human brain.

One of the ways that artificial intelligence is used to provide the most accurate images is through object recognition technology, which stores, synthesizes, and statistical data. Along with developing and refining algorithms and software tools, accurate object recognition technology also requires high-performance server systems that support numerous GPUs and object-oriented storage devices with significant capacity and quick access times.

</div><span class="text_page_counter">Trang 9</span><div class="page_container" data-page="9">

Furthermore, one of the drawbacks of this technology is the cost of paying for the investment in capacity packages, access speed, and computer hardware. In the future, object recognition technology will be able to identify objects smaller than those that can be seen by the human eye, such as microorganisms, vehicles, and people. Biotechnology industry bacteria support a thriving healthcare sector that safeguards people's health. The Yolov7 technology, created by Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao, is among the most well-liked object recognition systems available today.

2.1.3. Career opportunities

Today, AI is having a significant impact across several industries. There are several fields that are starting to apply and develop artificial intelligence, particularly object identification, which has been used in high-end apartments, warehouse management, human resource management, working time management, etc. There are still many untapped general uses for AI. Additionally, it offers a chance to advance and build object recognition AI.

Any field, though, must eventually develop AI and have practical features. In the area of object recognition, artificial intelligence has some of the following effects:

Vehicles: Identification of license plates in parking lots at businesses, schools, and enterprises.

Production: Control product lines on the conveyor belt, classify them, and assess worker performance.

Healthcare: Manage and monitor patients to determine when it is appropriate for them to start taking medication.

In many different industries, AI technology is constantly evolving. In addition to that, however, businesses in the technology sector need sufficient human resources. Following finishing a three-month internship and beginning employment there after the internship, an AI engineer can expect to make a minimum salary of 20 to 25 million VND, which is considered to be acceptable by enterprises. Companies like Viettle, Mobifone, and Maritime are the ones that are most welcoming to data scientists, artificial intelligence (AI), and data engineering in Vietnam.

2.2. IDENTIFY OBJECTS BY YOLOV7 TECHNOLOGY

</div><span class="text_page_counter">Trang 10</span><div class="page_container" data-page="10">

2.2.1. Introduction

Real-time object detection is a very important topic in computer vision, as it is often a necessary component in computer vision systems. The computing devices that execute real-time object detection is usually some mobile CPU or GPU, as well as various neural processing units (NPU) developed by major manufacturers.

In recent years, the real-time object detector is still developed for different edge device. For example, the development of MCUNet and NanoDet.

In this paper, we will present some of the new issues of this paper are summarized as follows:

- We design several trainable bag-of-freebies methods - The evolution of object detection methods

- We propose “extend” and “compound scaling” methods

- Effectively reduce about 40% parameters and 50% computation of state-of-the-art real-time object detector

- Has faster inference speed and higher detection accuracy

Figure 2.1: Comparison with other real-time object detectors

Unlike the previous YOLOv5 and YOLOv6, YOLOv7 comes from the author of YOLOv4 Alexey Bochkovskiy. Extended efficient layer aggregation networks (E-ELAN), model scaling, and a set of trainable bag-of-freebies are used.

</div><span class="text_page_counter">Trang 11</span><div class="page_container" data-page="11">

YOLOv7 outperforms all real-time Object Detection models, available at 30 FPS or more on GPU V100, in both speed and accuracy from 5 FPS to 160 FPS, and achieves the highest accuracy with 56.8% AP. YOLOv7-E6 (56 FPS on V100, 55.9% AP) outperforms high-end CNN backbones like ConvNeXt-XL Cascade-Mask R-CNN (8.6 FPS on A100, 55.2% AP) with 551% in speed and 0.7% in AP, as well as the Transformer home backbone SWIN-L Cascade-Mask R-CNN (9.2 FPS on And of course, in terms of speed, degree, and accuracy, YOLOv7 outperforms YOLOR, YOLX, Scaled-YOLOv4, YOLOv5, DETR, Deformable DETR, DINO-5scale-R50, ViT-Adapter-B, as well as many other Object Detection networks.

Additionally, YOLOv7 is trained entirely from scratch on COCO without the aid of any pretrained data.

(1) a faster and stronger network architecture; (2) a more effective feature integration method; (3) a more accurate detection method; (4) a more robust loss function;

(5) a more efficient label assignment method; (6) a more efficient training method. 2.2.2.2.Modelre-parameterization

Model re-parameterization techniques merge multiple computational modules into one at inference stage. The model re-parameterization technique can be regarded as an ensemble technique, and we can divide it into two categories, module-level ensemble and model-level ensemble.

There are two common practices for model-level re-parameterization:

- One is to train multiple identical models with different training data, and then average the weights of multiple trained models.

</div><span class="text_page_counter">Trang 12</span><div class="page_container" data-page="12">

- The other is to perform a weighted average of the weights of models at different iteration number.

This type of method splits a module into multiple identical or different module branches during training and integrates multiple branched modules into a completely equivalent module during inference

Model scaling is a method for increasing the size of the model for improved performance. With a scaling method that aggregates the depth, width, and resolution dimensions of the input image, model scaling is examined for the first time in EfficientNet.

Additionally, in order to achieve a good trade-off between the number of network parameters, computation, inference speed, and accuracy, the model scaling method typically uses various scaling factors, such as resolution (size of input image), depth (number of layer), width (number of channel), and stage (number of feature pyramid).

Figure 2.2: Compound scaling with EfficientNet 2.2.3. Architecture

2.2.3.1.Extendedefficientlayer aggregationnetworks

In most of the literature on designing, the main considerations are number of parameters, the amount of computation, and the computational density.

</div><span class="text_page_counter">Trang 13</span><div class="page_container" data-page="13">

Figure 2.3: Extended efficient layer aggregation networks

Instead of adjusting the gradient transmission path of the original architecture, the proposed extended ELAN (E-ELAN) uses group convolution to raise the cardinality of the added features and combines the features of various groups in a way that shuffles and merges attribute values. This mode of operation can improve the use of parameters and calculations as well as the features that are learned by various feature maps.

The architecture of CSPVoVNet, which is a variation of VoVNet and is depicted in Figure 2.3(b), also analyzes the gradient path, allowing the weights of various layers to learn more varied features.

Beside, A deeper network can efficiently learn and converge by controlling the shortest longest gradient path, according to ELAN in Figure 2.3(c). Finally, in this paper, we propose Extended-ELAN (E-ELAN) based on ELAN and its main architecture is shown in Figure 2.3 (d).

2.2.3.2.Model scaling forconcatenation-basedmodels

The main purpose of model scaling is to adjust some attributes of the model and generate models of different scales. The Scaling factors include in Resolution, Depth, Width and Stage.

Figure 2.4: Model scaling for concatenation-based models

</div><span class="text_page_counter">Trang 14</span><div class="page_container" data-page="14">

From (a) to (b), we mention that the output width of a computational block also grows when depth scaling is applied to concatenation-based models. The input width of the subsequent transmission layer will grow as a result of this phenomenon.

Therefore, we suggest (c), which states that only the depth in a computational block needs to be scaled when performing model scaling on concatenation-based models, with the remaining transmission layer being performed with corresponding width scaling.

2.2.4. Trainable bag-of-freebies

The trainable bag-of-freebies (BoF) are techniques added in training that can increase accuracy without increasing model processing time.

Model Level Ensemble performs a weighted average of the weights of models at different iteration number. Additionally, it has some function such as train multiple identical models with different training data, averages the weights of multiple trained models and splits and integrates the branched modules into a completely equivalent module.

Re-parameterization techniques involve averaging a set of model weights to create a model that is more robust to general patterns that it is trying to model.

Figure 2.5: Planned re-parameterized model

</div><span class="text_page_counter">Trang 15</span><div class="page_container" data-page="15">

A layer with residual or concatenation connections, its RepConv should not have an identity connection, according to our findings in the proposed planned re-parameterized model. In this case, RepConvN, which lacks identity connections, can take its place. In addition, RepConv actually combines 3×3 convolution, 1×1 convolution, and identity connection in one convolutional layer.

Deep supervision is a method that is frequently applied when deep networks are being trained. Its main idea is to increase the number of auxiliary heads in the middle layers of the network, while using assistant loss as a guide to weight the shallow layers of the network.

The model will be assisted in classifying and labeling objects to speed up high recognition speed if Auxiliary head is applied to all Classification or Semantic Segmentation lessons.

Figure 2.6: Coarse for auxiliary and fine for lead head label assigner. In contrast to the standard model (a), the schema in (b) has an auxiliary head. We suggest a lead head guided label assigner (d) and a coarse-to-fine lead head guided label assigner (e) in contrast to the conventional independent label assigner (c).

Furthermore, to obtain the labels of the training lead head and auxiliary head simultaneously, the proposed label assigner is optimized by lead head prediction and the ground truth. Details of the coarse-to-fine implementation method will be elaborated, as well as the constraint design.

Additionally, in order to apply the coarse for auxiliary and fine for lead losses method, we must pay attention to crucial factors such as Detection, Depth, Lead Head, GT, Assigner and Loss.

The main components in the optimization of labeling and prediction in this model include in The Lead Head and Aux Head (coarse-to-fine lead head guided). The lead head guided label assigner is primarily calculated using the lead head's prediction

</div>