Deep learning final project report car logo recognition

The study begins by reviewing the existing literature on image recognition, deep learning, and logo recognition to establish a foundation for the proposed system.. The application of mac

PHENIKAA UNIVERSITY

FACULTY OF ELECTRIC AND ELECTRICAL ENGINEERING

DEEP LEARNING FINAL PROJECT REPORT:

CAR LOGO RECOGNITION

STUDENTS:

Nguyen Ngoc Doan

Advisor: Dr Le Minh Huy

27/1/2024

2

Abstract With the increasing prevalence of automotive brands and the growing importance of visual identity

in marketing, the need for efficient car logo recognition systems has become crucial This research explores the development and implementation of an automated car logo recognition system using advanced computer vision techniques

The study begins by reviewing the existing literature on image recognition, deep learning, and logo recognition to establish a foundation for the proposed system Subsequently, a dataset comprising a diverse range of car logos is compiled and preprocessed to ensure the robustness of the model Various convolutional neural network (CNN) architectures, including state-of-the-art models, are evaluated to identify the most suitable one for the task

To address challenges such as varying lighting conditions, perspectives, and occlusions, the proposed system incorporates data augmentation, transfer learning, and fine-tuning strategies The training process involves optimizing hyperparameters and leveraging pre-trained models on large-scale datasets to enhance the model's generalization capabilities

The performance of the developed car logo recognition system is evaluated through rigorous testing on real-world images and compared with existing methods The metrics considered include accuracy, precision, recall, and F1 score Additionally, the system's computational efficiency and real-time applicability are assessed

The findings of this research contribute to the field of computer vision by providing insights into the challenges and opportunities associated with car logo recognition The proposed system offers

a practical solution for automakers, marketing analysts, and security systems, enabling efficient and reliable identification of car logos in diverse scenarios

Here is the link to my presentation video:https://youtu.be/yTIIyNejbXA

Link to the datasets:

https://drive.google.com/file/d/1TQQuT60bddyeGBVfwNOk6nxYavxQdZJD/view

3

4

Table of Contents

1 INTRODUCTION 5

2 DATASET 5

2.1 CAR LOGO DATASET 5

2.2 DATA COLLECTION 6

3 MODEL: 8

3.1 MODEL RCHITECTUREA 8

3.2 TRAINING AND VALUATIONE 9

4 CONCLUSION 12

REFERENCES 14

5

1 Introduction

Car logos are an important part of a car company's brand image Designed to create a first impression with customers, car logos can help increase brand awareness and capture the attention

of consumers

The classification of car logos is an important problem in the field of artificial intelligence When a new car company is launched, the classification of car logos will help car enthusiasts easily recognize and distinguish between different car manufacturers

In addition, the classification of car logos can also be applied in other areas such as advertising, marketing and market analysis For example, analyzing the number of cars by individual automakers on the street can provide important information to those interested in the car market

In this article, we will learn about artificial intelligence car logo identifying methods These methods can be used to build automatic car logo recognition systems and improve the efficiency

of human car logo recognition work

2 Dataset

2.1 Car logo dataset

Integrating car logos into a deep learning model introduces a fascinating dimension to the realm of artificial intelligence The application of machine learning algorithms to analyze and interpret 2D and 3D car logos allows for advanced recognition and classification capabilities

In the case of 2D logos, a deep learning model can be trained to identify these symbols across diverse media, such as printed materials, websites, and flat surfaces The simplicity of 2D representation lends itself well to machine learning algorithms, enabling accurate recognition and classification The model's ability to discern intricate details and variations in 2D logos contributes to a robust understanding of brand identity

Moving to 3D logos, the application of deep learning becomes even more intriguing Training a model to recognize and interpret the complexities of three-dimensional structures opens new possibilities in logo analysis This advanced capability could prove invaluable in scenarios such as augmented reality applications, where a deep learning model can discern and interact with 3D representations of car logos in real-time

6

The integration of car logos into a deep learning model not only enhances the efficiency of logo recognition but also lays the foundation for innovative applications in fields like autonomous vehicles, where quick and accurate logo identification can contribute to safe and efficient navigation Overall, the marriage of car logos and deep learning exemplifies the evolving synergy between traditional branding elements and cutting-edge artificial intelligence technologies

Figure 1 Example of Car logo

2.2 Data Collection

Firstly, clearly define the objectives of my car logo recognition project Specify the particular brands or types of logos you aim to recognize Build a list of specific automotive brands that will be the focus of my efforts Next, leverage web scraping tools or APIs to automatically download logo images from various sources such as official brand websites, automotive forums, and open datasets on platforms like Kaggle Download images either programmatically or manually from sources like press releases, media kits, and brand guidelines Ensure diversity in my data sources by including images captured under different lighting conditions, angles, and backgrounds Use filters to exclude irrelevant or low-quality images, and conduct quality control to eliminate duplicates, low-resolution images, or those with unclear logos Manual collection may be necessary, particularly when seeking high-quality images from reliable sources such as official brand websites To enhance dataset diversity, apply data augmentation techniques, incorporating random rotations, flips, and color adjustments Subsequently, pay attention to the labeling and annotation phase for each image

7

This step is crucial for supervised learning and model training Divide the dataset into training, validation, and testing sets This ensures proper evaluation and validation of the recognition model's performance Adhere to ethical standards and legal regulations during the collection and use of images Respect copyright and intellectual property rights associated with car logos Seek necessary permissions when required Document the details of the dataset, including the number of images per brand, specific challenges encountered, and any preprocessing steps applied Implement version control for the dataset to track changes, additions, and improvements over time By following these steps, I can create a comprehensive and well-organized dataset for car logo recognition This dataset serves as the foundation for training and evaluating machine learning models in this domain

Figure 2Brand Distribution

I employed Principal Component Analysis (PCA) as Figure 3 for data reduction and visualization The PCA-generated data was depicted in a two-dimensional (2D) plane In this 2D PCA representation, my dataset was projected onto two principal components, capturing the most significant variance while reducing dimensionality This transformation enabled us

to observe the distribution of data points on a simplified plane Consequently, it provided a nuanced view of the dataset's distribution, allowing exploration of relationships in three dimensions The 2D visualization yielded valuable insights; notably, the data did not exhibit clear clustering, as samples from each class showed overlapping patterns with others

8

Figure 3 Visualization of PCA

3 Model:

3.1 Model Architecture

Figure 4 Model Architecture

• Input Layer: The model begins with an input layer The input shape is expected to have

a shape of (300, 200, 3), indicating a 3-channel image with dimensions 300x200 pixels

• Convolutional Layer 1: Convolutional layer with 32 filters, a kernel size of unspecified dimensions, and a Rectified Linear Unit (ReLU) activation function

( ) max 0,( )

RELU z = (3.1)

• Pool: Max pooling layer with a pool size of (2, 2), reducing the spatial dimensions by half

9

• Convolutional Layer 2: Another convolutional layer with 64 filters and ReLU activation

• Pool: Max pooling layer with a pool size of (2, 2), reducing the spatial dimensions by half

• Dropout Layer : Dropout layer with a dropout rate of 0.5, aimed at preventing overfitting

• Output: Two Dense layers are included The first has 256 units with ReLU activation, and the second has a number of units equal to the number of classes with softmax activation, suitable for multi-class classification

𝑠𝑜𝑓𝑡𝑚𝑎𝑥 𝑧( 𝑖) =∑𝑘𝑒𝑧𝑖𝑒𝑧𝑗

𝑗=1 (3.2) Based on the presented model, several conclusions can be drawn regarding its suitability for classifying car logos The model employs a Convolutional Neural Network (CNN) architecture, well-suited for image processing and feature extraction from car logos Dropout layers are used after Conv2D layers and before the Dense layer to mitigate the risk of overfitting, a sound strategy for ensuring model generalization ReLU activation is used in both Conv2D and Dense layers to facilitate learning of non-linear features and address the vanishing gradient problem The final layer uses softmax activation, fitting for multi-class classification problems It transforms the output into a probability distribution across classes, facilitating easy interpretation of class probabilities In this model, the chosen loss function is categorical_crossentropy, suitable for multi-class classification with softmax activation It optimizes the reduction of information loss between predicted and actual probability distributions

𝐿𝑜𝑠𝑠 = − ∑𝑜𝑢𝑝𝑢𝑡 𝑠𝑖𝑧𝑒𝑦𝑖 × log𝑦𝑖

𝑖=1 (3.3)

3.2 Training and Evaluation

10

Figure 5 Training 1

With 10 epochs I got this result, it was overfitting so I changed a lot of weight

Figure 6.Training 2 With 50 epochs and after changing the weights, I see more overfitting

11

Figure 7.Training Result

I have been looking at several ways to reduce overfitting Firstly, adding Dropout layers after Dense layers is a common practice Dropout helps prevent over-reliance on specific neurons by randomly deactivating them during training, introducing a level of robustness.Regularization techniques, such as L1 or L2 regularization, can be applied to penalize large weights in Dense layers This discourages the model from fitting noise in the training data and promotes a more generalized representation.Reducing the size of the model, either by decreasing the number of neurons or layers, is another effective approach A simpler model is less prone to overfitting, especially when the dataset is not extensive Data augmentation is valuable for generating additional training samples by applying random transformations to the existing data This helps the model become more invariant to variations

in the input data.Early stopping is a practical strategy to halt training when the model's performance on the validation set plateaus, preventing it from overfitting to the training data Finally, adjusting the learning rate during training using a schedule can be beneficial This allows for fine-tuning the rate of parameter updates, potentially improving convergence and reducing overfitting

12

Figure 8.Confusion Matrix

Figure 9 Test with some brands

4 Conclusion

In conclusion, the implementation of car logo recognition using the specified model involves

a series of thoughtful considerations and adjustments The convolutional neural network (CNN) architecture, comprising convolutional layers, pooling layers, and densely connected layers, serves

13

as a robust framework for the task To address the challenge of overfitting, various strategies, including Dropout layers, regularization, and data augmentation, have been applied

The model's ability to recognize car logos relies on its capacity to extract meaningful features from the input images The convolutional layers play a crucial role in capturing hierarchical features, while the subsequent densely connected layers facilitate the learning of complex patterns and representations The incorporation of Dropout layers contributes to preventing overfitting, enhancing the model's generalization capability to unseen data

Data augmentation, with random transformations applied to the training dataset, enriches the diversity of input samples, making the model more resilient to variations and improving its robustness The utilization of regularization techniques further fine-tunes the model's weights, discouraging it from fitting noise in the training data

Continuous monitoring of the model's performance, as exemplified by early stopping and dynamic learning rate schedules, ensures that the training process adapts to the data's characteristics The convergence of these strategies in the specified model exemplifies a comprehensive approach to car logo recognition, where the model's ability to generalize to diverse logos and variations is paramount

In summary, the car logo recognition model, with its carefully designed architecture and mitigating strategies against overfitting, presents a powerful tool for accurately identifying and classifying car logos in diverse scenarios The success of such a model hinges on the interplay of architectural choices and optimization techniques, showcasing the continuous refinement and adaptability required in the field of deep learning for image recognition tasks

14

References [1] Deep Learning Course - Dr Minh Huy Le

[2] Cars logo recognition by using of backpropagation neural networks - Jabbar Majeed Sadeq

a, Brzo Aziz Qadir b, Hayder Hassan Abbas

[3] Vehicle Logo Recognition with an Ensemble of Classifiers - Bogusław Cyganek, Michał Woźniak