Vision Quest Challenge

Inspiration

Our inspiration stemmed from addressing the challenge of class imbalance in image datasets, which often leads to biased models. We aimed to leverage transfer learning and hybrid sampling techniques to tackle this issue and improve classification accuracy.

What it does

Our project focuses on enhancing the accuracy of image classification tasks by addressing class imbalance through a combination of transfer learning, feature extraction using a pre-trained model (ResNet50), and hybrid sampling techniques (SMOTEENN). The end result is a robust classification model capable of achieving high accuracy even with imbalanced datasets.

How we built it

Data Preparation: We loaded the image dataset along with class labels and preprocessed it to make it compatible with the ResNet50 pre-trained model. Transfer Learning: Utilizing the ResNet50 model, we performed feature extraction from the images and added a Dense layer for classification. Handling Class Imbalance: Recognizing the imbalance in the dataset, we applied the SMOTEENN hybrid sampling method to balance the class distribution, combining oversampling and undersampling techniques. Model Training: With the balanced dataset, we trained an XGBoost classifier to classify images, achieving a high testing accuracy score.

Challenges we ran into

Class Imbalance: Dealing with class imbalance presented a significant challenge, as it could lead to biased model predictions if not addressed properly. Model Optimization: Fine-tuning the parameters of the XGBoost classifier and optimizing the hybrid sampling technique required experimentation and iteration. Computational Resources: Training deep learning models and performing sampling techniques on large datasets demanded substantial computational resources.

Accomplishments that we're proud of

Effective Class Balancing: Successfully implementing the SMOTEENN hybrid sampling method helped in mitigating the effects of class imbalance, resulting in a more robust and fair model. High Accuracy: Achieving a testing accuracy score of almost 95.2% demonstrates the effectiveness of our approach in improving classification performance. Utilizing Transfer Learning: Leveraging transfer learning with the ResNet50 model allowed us to extract meaningful features from images and build upon existing knowledge.

What we learned

Importance of Class Balance: Addressing class imbalance is crucial for building accurate and fair machine learning models, especially in image classification tasks. Hybrid Sampling Techniques: Exploring and implementing hybrid sampling techniques like SMOTEENN provided insights into effectively handling imbalanced datasets. Transfer Learning Benefits: Leveraging pre-trained models for feature extraction can significantly streamline model development and improve performance, especially in scenarios with limited data.

What's next for Vision Quest Challenge

Moving forward, we aim to further enhance the robustness and generalization of our model by: Exploring additional transfer learning architectures and pre-trained models for feature extraction. Investigating advanced sampling techniques and model ensembles to improve classification accuracy and resilience to class imbalance. Extending the project to handle real-time image classification tasks and deploying it in practical applications such as healthcare diagnostics or environmental monitoring.