This project explores deep learning techniques for binary image classification on a low-resource dataset. Specifically, we utilized the Beans dataset, consisting of images of healthy and unhealthy bean plant leaves, as a proxy for screening rare genetic disorders. The insights gained from this task can inform the development of more accessible and cost-effective diagnostic tools. Dataset: https://www.kaggle.com/datasets/csafrit2/plant-leaves-for-image-classification

Model Architecture: Implemented a custom version of AlexNet due to its simplicity and proven effectiveness. Hyperparameters:

• Learning Rate: 0.01
• Batch Size: 16
• Epochs: 50
• Optimizer: Stochastic Gradient Descent
• Loss: Cross Entropy Loss
• Performance: Trained on the Beans dataset, achieving notable results on the validation set. Detailed performance is presented in the confusion matrix.

Model Architecture: A CNN-based autoencoder with distinct encoder and decoder components.

• Encoder: 4 convolutional layers with ReLU activations and batch normalization, followed by 2 linear layers.
• Decoder: 4 convolutional layers with ReLU activations and batch normalization, followed by 2 linear layers.
• Classifier: Uses the encoder layer, followed by 3 linear layers for classification. Hyperparameters:
• Encoder/Decoder Learning Rate: 0.01
• Classifier Learning Rate: 0.001
• Batch Size: 4 (Autoencoder), 16 (Classifier)
• Epochs: 7 (Autoencoder), 30 (Classifier)
• Optimizer: Adam (Autoencoder), SGD (Classifier)
• Loss: Mean Squared Error (Autoencoder), Cross Entropy Loss (Classifier)
• Performance: Evaluated on the Beans dataset with results detailed in a confusion matrix.

Discussed various approaches for effective classification with limited labeled samples, including data augmentation, transfer learning, shallow networks, active learning, and ensemble methods.

• CNN Accuracy: Achieved promising results, with learning curves and a confusion matrix for the validation set.
• Autoencoder Performance: Demonstrated effective image reconstruction and classification capabilities.

• Code: Python scripts for CNN and Autoencoder implementations.
• Models: Saved model parameters and checkpoints.
• Results: Confusion matrices, learning curves, and performance metrics.
• Blind Test Predictions: Results for the blind test dataset in CSV format. This project showcases the application of deep learning models to low-resource datasets, highlighting the potential for developing scalable and cost-effective diagnostic tools in various fields.

For more details, read the Report pdf file. Thanks :)