The research delved into multiple facets of advanced computer vision, exploring activation functions, CNN refinement for CIFAR10, transfer learning, fine-tuning, and object detection. In assessing activation functions, ReLU emerged as dominant, while optimal learning rates and network complexity significantly influenced convergence. The adapted CNN architecture for CIFAR10 remarkably achieved 82% across accuracy, precision, recall, and F1-Score, efficiently leveraging 531,818 parameters. Notable strategies included expanded layers, diverse activation functions, and optimization techniques. Transfer learning and fine-tuning highlighted the importance of learning rates and epochs, with AdaMax and Adam optimizers excelling. Model architecture evaluation favored MobileNetV2 for image classification. Object detection comparisons identified YOLOv8 as the most proficient, excelling in precision and recall.

This research delves into the exploration of activation functions within the neural network framework. The study revealed ReLU as the predominant performer across various datasets, showcasing its efficacy in enabling convergence. Notably, learning rates between 0.1 and 1 were identified as optimal for achieving convergence on the Circle dataset, while extremes led to output oscillations. Additionally, variations in network layers and neurons influenced decision boundary complexity and convergence speed, highlighting the crucial balance needed in model complexity. Evaluation of optimization algorithms highlighted Adadelta's superiority, exhibiting higher training and testing accuracy, along with reduced loss in comparison to other methods.

The study focused on refining TensorFlow's CNN tutorial for improved performance on the CIFAR10 dataset. CIFAR10 comprises 60,000 color images categorized into 10 classes, each containing 6,000 images. This dataset is split into 50,000 training images and 10,000 testing images. Each image is a 32x32 pixel array, making it a challenging yet standard benchmark in the field of computer vision due to its diverse classes and low resolution.

The adapted CNN architecture achieved remarkable enhancements, attaining an outstanding 82% Accuracy, Precision, Recall, and F1-Score on the CIFAR10 test set. These improvements were made while judiciously utilizing 531,818 model parameters, showcasing the efficiency and optimization achieved through the adapted architecture.

The architectural enhancements focused on expanding convolutional and dense layers, incorporating optimal activation functions like softmax, ELU, and ReLU, and employing optimization techniques such as the Adam optimizer with gradient clipping. Additional strategies included integrating dropout and batch normalization layers, utilizing a larger batch size and epochs, and incorporating an early stopping callback.

Observations from the enhanced model revealed notable outcomes: training and validation curves displayed gradual convergence, indicating minimized overfitting. However, challenges in accurately classifying specific classes, notably cat and dog, were identified from the confusion matrix. Despite this, a distinct diagonal line in the matrix highlighted the model's generally accurate classification across most classes.

The proposed architecture was further refined to achieve a more efficient model. The adapted architecture, named CIFAR DecaLuminarNet, achieved an impressive 87% Accuracy, Precision, Recall, and F1-Score on the CIFAR10 test set, utilizing 9,136,778 parameters. This model was trained for 150 epochs with a batch size of 512. The DecaLuminarNet conveys deep and illuminating insights into the model's architecture, with Deca representing the 10 classes in CIFAR10 and LuminarNet signifying the model's illuminating performance over the first proposed architecture.

The refined architecture focused on expanding convolutional layers, incorporating renowned optimal activation functions like softmax and ELU, and employing optimization techniques such as the Adam optimizer with gradient clipping. Additional strategies included integrating dropout and batch normalization layers, utilizing a larger batch size and epochs, and incorporating an early stopping and reduce learning rate on plateau callbacks.

Observations from the enhanced model revealed notable outcomes: training and validation curves displayed gradual convergence, indicating minimized overfitting. However, challenges in accurately classifying specific classes, notably cat and dog persisted. Despite this, a distinct diagonal line in the matrix highlighted the model's generally accurate classification across most classes.

This research explored advanced computer vision techniques, including transfer learning, fine-tuning, learning rates, optimizers, and model architectures, aiming to enhance image classification. The study aimed to identify crucial elements in the classification process and assess the combined impact of these techniques on overall performance.

Smaller learning rates, particularly 0.001, demonstrated superior initial performance with the Cats and Dogs dataset, resulting in higher validation accuracy and lower loss. Extremes (0.1 and 0.0001) led to poorer outcomes. Increasing the number of epochs notably improved model convergence and validation accuracy.

AdaMax and Adam optimizers showcased superior performance due to their adaptive learning rate strategies, resulting in higher accuracy and reduced loss. Conversely, AdaDelta's limitations in capturing intricate dataset patterns led to inferior performance. Additionally, repeated performance jumps during fine-tuning suggested instability in model performance.

Various model architectures were evaluated for image classification. MobileNetV2 displayed strong initial validation accuracy (Val Acc) of 0.9678 and excelled in fine-tuning with a Val Acc of 0.9913, achieving the highest Test Acc of 0.9948. ResNet50 and EfficientNetB0 emerged as noteworthy competitors, showcasing high accuracy and low loss, while InceptionV3 displayed the lowest validation loss.

1. Static Params: Loss = Cross Entropy, LR = 0.001, Initial Epochs = 10, Fine-Tune Epochs = 10, Base Model = MobileNetV2
2. Static Params: Loss = Cross Entropy, LR = 0.001, Initial Epochs = 10, Fine-Tune Epochs = 10
3. Static Params: Loss = Cross Entropy, LR = 0.001, Initial Epochs = 10, Fine-Tune Epochs = 10

In this project, object classification models like ResNet50, VGG16, and MobileNet were employed to classify images into two categories: those containing pizza and those that didn't. Subsequently, an annotated dataset encompassing diverse pizza types and ingredients was created using RoboFlow. This dataset was then used to train object-detection models—YOLOv8, YOLOv5, RetinaNet—aimed at identifying pizzas and their ingredients. Loss and accuracy graphs were generated for evaluation. Although a basic implementation of DETR was explored, it was not included in the final assessment. The culmination of these techniques led to the construction of a comprehensive pizza analysis system.

The code for this research was executed on Google Colab, with provided environment files (environment.yml and requirements.txt) for local execution of object detectors. The main directory structure includes the following folders:

• Papers: Contains references to any papers cited in the documentation.
• Part_1_Object_Classification: Includes files pertinent to the Object Classification task.
• Part_2_Building a Dataset: Encompasses files related to the creation of an annotated dataset.
• Part_3_Object_Detection: Holds files related to the Object Detection task.
• Part_3_Results: Consists of diagrams or images created for evaluation of both object classification and object detection tasks.
• pizza_data: Contains the Kaggle Pizza Data Dataset [1].
• pizza_classification: Holds the Kaggle Pizza Classification Dataset [2].

Object classification is a core task in computer vision, involving the detection and categorization of objects based on predefined classes. Models trained on labeled datasets can assign labels to images with varying confidence levels, often requiring a threshold to filter low-confidence predictions.

ImageNet (2009) is a widely used dataset for training computer vision models, though not directly utilized in this project. Pre-trained models with ImageNet weights were employed.

Introduced in 2014, VGG16 excels in image classification despite its relative simplicity compared to newer architectures, focusing on capturing intricate patterns within images.

ResNet50 (2015) addresses deep network training issues using residual learning blocks and skip connections, enhancing accuracy in image recognition by overcoming the vanishing gradient problem.

Designed in 2017 for mobile and edge devices with limited computational resources, MobileNet employs depth-wise separable convolutions, striking a balance between accuracy and efficiency for real-time applications.

Three classification models—ResNet50, VGG16, and MobileNet—were chosen for implementation in this study. Apart from their popularity, these models were selected due to their manageable computational requirements and diverse architectures. A Kaggle-retrieved dataset [2] was used for comparison purposes. Images were resized to 224x224 to accommodate model input size requirements, ensuring fair evaluation among the architectures.

Utilizing pre-trained weights from ImageNet, the classification process commenced by feeding images to each model. The predict function returned lists of tuples containing label identifiers, names, and associated confidence scores for each image. Top five labels per image, ranked by confidence, were displayed and saved into JSON files corresponding to each model used.

Despite their capability to classify various labels, these models were limited to classifying a pizza without distinguishing its toppings, as none were part of the known label list.

Building a robust annotated dataset is a time-consuming process vital for object detection. It begins with dataset selection and label definition before the annotation process. Bounding boxes were used for this project, while other tasks might employ polygon annotation. Precise and consistent annotations are essential, facilitated by tools like Roboflow.

Established in 2019, Roboflow is a user-friendly platform offering annotation, augmentation, and organization tools, aiding efficient dataset management.

For object detection, a Kaggle-sourced pizza dataset of around 9000 images was utilized [1]. Approximately 1500 images were annotated, focusing on 16 common pizza ingredients for labeling.

Chosen ingredient labels:

1. Arugula
2. Bacon
3. Basil
4. Broccoli
5. Cheese
6. Chicken
7. Corn
8. Ham
9. Mushroom
10. Olives
11. Onion
12. Pepperoni
13. Peppers
14. Pineapple
15. Pizza
16. Tomatoes

To streamline annotation and system performance, certain ingredients were omitted. Roboflow's tools were used for label management, bounding box creation, and image sorting.

Following annotation, the dataset was split into training, validation, and testing sets in a 60%-20%-20% ratio. Augmentations, like rotation and blur, were exclusively applied to the training set, resulting in 2544, 284, and 283 images, respectively. The dataset was exported in various formats via Roboflow to suit object detection model requirements.

This meticulous dataset preparation laid the foundation for subsequent training and evaluation phases of the object detection model.

Roboflow introduced its own model (Roboflow 3.0) utilizing the labeled dataset mentioned earlier for training. The dataset is available at https://app.roboflow.com/advanced-computer-vision-assignment/pizza-object-detector/deploy/7, while the resultant model can be accessed via the QR code or through the deployment dashboard displayed below.

The QR code can be scanned to access the deployment dashboard below:

To cite the dataset, please use the following BibTeX format:

@misc{ pizza-object-detector_dataset,
    title = { Pizza Object Detector Dataset },
    type = { Open Source Dataset },
    author = {Matthias Bartolo and Jerome Agius and Isaac Muscat},
    howpublished = { \url{ https://universe.roboflow.com/advanced-computer-vision-assignment/pizza-object-detector } },
    url = { https://universe.roboflow.com/advanced-computer-vision-assignment/pizza-object-detector },
    journal = { Roboflow Universe },
    publisher = { Roboflow },
    year = { 2023 },
    month = { nov },
}

Object Detection in computer vision involves identifying objects and their locations within images or videos, incorporating bounding boxes and class labels. Unlike object classification, it pinpoints an object's position through bounding boxes and assigns appropriate class labels.

RetinaNet (2017) addresses class imbalance and localization with its Focal Loss, emphasizing difficult examples during training. Its feature pyramid network handles multi-scale features, making it proficient in complex, multi-scale object detection scenarios.

YOLOv5 (2020) by Ultralytics improves YOLO's legacy with a more streamlined architecture, offering customization and multi-platform deployment. Variants like YOLOv5x and YOLOv5s cater to different computational requirements, focusing on simplicity, efficiency, and performance.

YOLOv8 (2023) continues advancements, utilizing CSPDarknet53 for efficient image processing while maintaining high accuracy. Variants like YOLOv8-CSP and YOLOv8-Darknet provide options for different computational resources and use cases.

DETR (2020) from Facebook AI Research replaces traditional anchor-based methods with transformers for end-to-end object detection. It handles multiple objects effectively in complex environments, especially with small-sized objects.

The project implemented YOLOv5, YOLOv8, RetinaNet, and DETR for object detection, leveraging their specific strengths. Each model underwent dataset preparation, training, and evaluation. Post-training, examples of object detection in test images were showcased to assess model performance and capabilities.

The evaluation compared YOLOv5, YOLOv8, RetinaNet, and DETR models. YOLOv8 showcased superior overall performance compared to YOLOv5, demonstrating better precision and recall in most scenarios despite the use of early stopping. While YOLOv5 displayed more consistent gradient reduction and higher precision at greater recall levels in precision-recall curves, YOLOv8 emerged as the best-performing model considering multiple evaluation metrics and curves. RetinaNet, an older architecture, showed lower precision but higher recall, producing acceptable results compared to modern YOLO architectures. Though DETR was implemented without specific graphs, a table with Average Precision and Recall values was provided, enabling comparison against other models. In summary, while YOLOv5 excelled in certain aspects, the comprehensive evaluation favored YOLOv8 as the top-performing model in this evaluation.

RetinaNet MAP Graph

RetinaNet PR Curve

YOLOv5 Confusion Matrix

YOLOv8 Confusion Matrix

The evaluation highlighted VGG16 as the top-performing image classifier, achieving the highest precision and F1-Score among the implemented models. For object detection tasks on a custom annotated pizza dataset, YOLOv8 emerged as the best-performing model, exhibiting superior curves compared to others. This study underscores the performance of various models in image classification and object detection tasks using a custom annotated dataset. While certain models excelled in this specific dataset, results could have varied with a different dataset selection. Additionally, the choice of pre-defined labels might have impacted the performance of object detection models.

[1] M. Bryant, Pizza images with topping labels, https://www.kaggle.com/datasets/michaelbryantds/pizza-images-with-topping- labels/, Jun. 2019.

[2] Project_SHS, Pizza images with topping labels, https://www.kaggle.com/datasets/projectshs/pizza-classification-data/, Dec. 2022.