This project focuses on image classification using Convolutional Neural Networks (CNNs) for object recognition.
A custom CNN architecture named MetENet was designed, trained and evaluated on the CIFAR-10 dataset.
The project explores the complete machine-learning workflow, including:
- dataset analysis and preprocessing,
- CNN architecture design,
- model optimization,
- training and validation,
- performance evaluation,
- cross-validation,
- comparison with well-known CNN architectures,
- and generalization on additional datasets.
The objective was to develop a CNN capable of achieving competitive classification performance while maintaining a relatively efficient architecture.
This work was developed as an academic integrative project during the 2024–2025 academic year.
Developed by:
Eya Sahli & Montassar Laboudi
Supervisor:
Ahcen Aliouet
The project focuses on:
- Deep Learning
- Computer Vision
- Image Classification
- Convolutional Neural Networks
- Model Optimization
- Performance Evaluation
The main dataset used in this project is CIFAR-10.
CIFAR-10 contains 60,000 RGB images of size 32 × 32 pixels, distributed across 10 object categories:
- Airplane
- Automobile
- Bird
- Cat
- Deer
- Dog
- Frog
- Horse
- Ship
- Truck
The dataset is balanced across the different classes.
Two validation strategies were evaluated.
The dataset was divided as follows:
- 70% Training
- 10% Validation
- 20% Testing
A second evaluation was performed using 5-fold cross-validation.
The model was trained five times, with four folds used for training and one fold used for validation during each iteration.
The resulting accuracies were:
| Fold | Accuracy |
|---|---|
| Fold 1 | 86.64% |
| Fold 2 | 86.64% |
| Fold 3 | 87.05% |
| Fold 4 | 86.32% |
| Fold 5 | 86.44% |
| Average | 86.62% |
The standard dataset split achieved an accuracy of 86.55%, showing very similar performance between both evaluation strategies.
MetENet is a custom CNN architecture inspired by the principles of deeper convolutional networks such as VGG.
Several architectures were explored during development before selecting a four-block convolutional architecture.
Conv 3×3 — 32 filters
Conv 3×3 — 32 filters
Batch Normalization
ReLU
Max Pooling 2×2
Conv 3×3 — 64 filters
Conv 3×3 — 64 filters
Batch Normalization
ReLU
Max Pooling 2×2
Conv 3×3 — 128 filters
Conv 3×3 — 128 filters
Batch Normalization
ReLU
Max Pooling 2×2
Conv 3×3 — 256 filters
Conv 3×3 — 256 filters
Batch Normalization
ReLU
Max Pooling 2×2
The progressive increase in the number of filters allows the network to learn increasingly complex image representations.
Early layers mainly detect low-level characteristics such as edges and textures, while deeper layers capture more complex visual patterns.
Dropout was introduced to reduce overfitting.
Different dropout rates were used depending on the depth of the network:
| Layer | Dropout |
|---|---|
| Block 1 | 0.10 |
| Block 2 | 0.20 |
| Block 3 | 0.30 |
| Block 4 | 0.30 |
| Fully Connected Layer | 0.35 |
This progressive regularization strategy helps preserve low-level feature learning in the first layers while applying stronger regularization to deeper representations.
The network was trained using the Adam optimizer.
Several parameters were adjusted during the optimization phase, including:
- learning rate,
- number of epochs,
- mini-batch size,
- validation frequency,
- dropout rate,
- number of convolutional blocks,
- number of filters.
An early stopping mechanism was also used to reduce unnecessary training once validation performance stopped improving.
The selected MetENet configuration reached approximately 86.55% classification accuracy on CIFAR-10.
The training curves show progressive improvement in classification accuracy together with a reduction in the loss function.
The confusion matrix provides a detailed view of the model's classification behavior across all CIFAR-10 classes.
Strong classification performance can be observed for classes such as:
- automobile,
- airplane,
- frog,
- horse,
- ship,
- truck.
Some confusion remains between visually similar categories such as cat and dog, which represents a common challenge on CIFAR-10.
The following examples show predictions generated by MetENet together with their confidence probabilities.
MetENet was compared with several widely used CNN architectures.
| Model | Accuracy | F1-Score | Error Rate | Training Time |
|---|---|---|---|---|
| AlexNet | 80.07% | 80.07% | 19.93% | 14 min 11 s |
| SqueezeNet | 85.79% | 85.79% | 14.21% | 14 min 53 s |
| MetENet | 86.55% | 86.55% | 13.45% | 11 min 55 s |
| GoogLeNet | 87.79% | 87.79% | 12.21% | 31 min 53 s |
| MobileNetV2 | 90.91% | 90.91% | 9.09% | 364 min 11 s |
| ResNet18 | 91.80% | 91.80% | 8.20% | 17 min 59 s |
The results show that MetENet provides competitive performance while maintaining a relatively efficient training time in the experimental environment.
A second comparison was performed by training selected architectures entirely from scratch.
| Model | Accuracy |
|---|---|
| AlexNet | 76.79% |
| SqueezeNet | 83.36% |
| GoogLeNet | 84.20% |
| MetENet | 86.55% |
| ResNet18 | 90.68% |
In this experiment, MetENet achieved higher accuracy than AlexNet, SqueezeNet and GoogLeNet.
To evaluate whether the architecture could generalize beyond CIFAR-10, MetENet was also evaluated on additional image datasets.
Fashion-MNIST contains 70,000 grayscale images of clothing items distributed across 10 classes.
MetENet achieved:
Example predictions:
These results show that the architecture is capable of extracting useful features from a dataset with visual characteristics significantly different from CIFAR-10.
The SVHN dataset contains real-world images of house numbers captured from Google Street View.
Compared with CIFAR-10 and Fashion-MNIST, the images may contain additional challenges such as:
- blur,
- illumination variations,
- real-world backgrounds,
- class imbalance.
MetENet achieved:
Example predictions:
The results demonstrate the ability of the model to generalize to a different image-recognition problem involving real-world digit images.
| Experiment | Accuracy |
|---|---|
| CIFAR-10 — Standard Split | 86.55% |
| CIFAR-10 — 5-Fold Cross-Validation | 86.62% |
| Fashion-MNIST | 92.72% |
| SVHN | 95.3941% |
| Area | Technologies |
|---|---|
| Deep Learning | Convolutional Neural Networks |
| Development | MATLAB |
| Optimization | Adam |
| Validation | Train/Validation/Test, 5-Fold Cross-Validation |
| Evaluation | Accuracy, Precision, Recall, F1-Score, Confusion Matrix |
| Computer Vision | Image Classification, Object Recognition |
| Datasets | CIFAR-10, Fashion-MNIST, SVHN |
| Hardware | NVIDIA GPU |
- Deep Learning
- Convolutional Neural Networks
- Computer Vision
- Image Classification
- CNN Architecture Design
- Hyperparameter Optimization
- Cross-Validation
- Model Evaluation
- Confusion Matrix Analysis
- MATLAB
- Experimental Analysis
- Scientific Documentation
The project resulted in the development of MetENet, a custom CNN architecture designed for image classification.
The model achieved 86.55% accuracy on CIFAR-10 and showed consistent performance through 5-fold cross-validation with an average accuracy of 86.62%.
The comparison with established CNN architectures showed competitive performance, particularly when considering training efficiency.
The additional experiments on Fashion-MNIST (92.72%) and SVHN (95.3941%) demonstrated that the architecture can also generalize to datasets with substantially different visual characteristics.
Possible extensions of this work include:
- additional architecture optimization,
- data augmentation strategies,
- automated hyperparameter optimization,
- lightweight deployment on embedded platforms,
- model compression and quantization,
- real-time image classification,
- evaluation on additional real-world datasets.
This repository contains the MATLAB implementation of the MetENet project together with selected experimental results and visualizations.
MetENet-CNN-Image-Classification-MATLAB/
│
├── src/
│ └── MATLAB source code
│
├── figures/
│ ├── training-progress.png
│ ├── cifar10-confusion-matrix.png
│ ├── cifar10-predictions.png
│ ├── fashion-mnist-predictions.png
│ └── svhn-predictions.png
│
└── README.md
Eya Sahli