CNN Architectures for Image Classification

Author

Arpit Aggarwal

Introduction to the Project

In this project, different CNN Architectures like AlexNet, VGG-16, VGG-19, InceptionNet, DenseNet-121 and ResNet-50 were used for the task of Dog-Cat image classification. The input to the CNN networks was a (224 x 224 x 3) image and the number of classes were 2, where '0' was for a cat and '1' was for a dog. The CNN architectures were implemented in PyTorch and the loss function was Cross Entropy Loss. The hyperparameters to be tuned were: Number of epochs(e), Learning Rate(lr), momentum(m), weight decay(wd) and batch size(bs).

Data

The data for the task of Dog-Cat image classification can be downloaded from: https://drive.google.com/drive/folders/1EdVqRCT1NSYT6Ge-SvAIu7R5i9Og2tiO?usp=sharing. The dataset has been divided into three sets: Training data, Validation data and Testing data. The analysis of different CNN architectures for Dog-Cat image classification was done on comparing the Training Accuracy and Validation Accuracy values.

Methods for preventing overfitting

The methods that were applied to prevent overfitting were: Label smoothing, Weight Decay, Data Augmentation and Dropout. To learn more about label smoothing regularization technique, refer to the link provided in the Credits section.

Results

The results after using different CNN architectures are given below:

1. AlexNet
   

Not Pre-trained: Training Accuracy = 96.64% and Validation Accuracy = 93.59% (e = 50, lr = 0.005, m = 0.9, bs = 64, wd = 0.001)

Pre-trained: Training Accuracy = 98.01% and Validation Accuracy = 96.63% (e = 50, lr = 0.005, m = 0.9, bs = 64, wd = 0.001)

2. VGG-16
   

Not Pre-trained: Training Accuracy = 98.32% and Validation Accuracy = 97.13% (e = 100, lr = 0.005, m = 0.9, bs = 32, wd = 5e-4)

Pre-trained: Training Accuracy = 99.22% and Validation Accuracy = 98.58% (e = 30, lr = 0.001, m = 0.9, bs = 32, wd = 5e-4)

3. VGG-19
   

Not Pre-trained: Training Accuracy = 98.71% and Validation Accuracy = 97.05% (e = 100, lr = 0.005, m = 0.9, bs = 32, wd = 5e-4)

Pre-trained: Training Accuracy = 99.15% and Validation Accuracy = 98.53% (e = 30, lr = 0.001, m = 0.9, bs = 32, wd = 5e-4)

4. DenseNet-121
   

Not Pre-trained: Training Accuracy = 99.2% and Validation Accuracy = 92.9% (e = 60, lr = 0.003, m = 0.9, bs = 32, wd = 0.001)

Pre-trained: Training Accuracy = 98.73% and Validation Accuracy = 98.01% (e = 40, lr = 0.01, m = 0.9, bs = 32, wd = 5e-4)

5. InceptionNet
   

Not Pre-trained: Training Accuracy = 98.1% and Validation Accuracy = 93.8% (e = 60, lr = 0.003, m = 0.9, bs = 32, wd = 0.001)

Pre-trained: Training Accuracy = 99.94% and Validation Accuracy = 99.13% (e = 40, lr = 0.005, m = 0.9, bs = 32, wd = 5e-4)

6. ResNet-50
   

Not Pre-trained: Training Accuracy = 96.33% and Validation Accuracy = 92.72% (e = 100, lr = 1e-3, m = 0.9, bs = 32, wd = 5e-4)

Pre-trained: Training Accuracy = 99.43% and Validation Accuracy = 98.43% (e = 30, lr = 0.005, m = 0.9, bs = 32, wd = 5e-4)

Software Required

To run the jupyter notebooks, use Python 3. Standard libraries like Numpy and PyTorch are used.

Credits

The following links were helpful for this project:

1. https://www.youtube.com/channel/UC88RC_4egFjV9jfjBHwDuvg
2. https://github.com/pytorch/tutorials
3. https://towardsdatascience.com/what-is-label-smoothing-108debd7ef06
4. https://leimao.github.io/blog/Label-Smoothing/