This repository contains code for a flame classification model that can classify images of flames into three categories: ethanol, pentane, and propanol.
The dataset is organized into train, validation, and test sets with a split ratio of 70%, 15%, and 15% respectively. Images are preprocessed with normalization and augmentation techniques such as rotations, flips, and scaling.
The model architecture is based on the ResNet-34 convolutional neural network. The final fully connected layer is modified to output predictions for the three flame categories.
The model is trained using the Adam optimizer with a learning rate of 1e-4 for 20 epochs. During training, the model's performance is evaluated on both the training and validation sets. Adjusted layer freezing and learning rate experiments are conducted for potential performance improvements.
After training, the model achieves an accuracy of 80.47% on the validation set.
Visualization of intermediate layers of the model is provided for better understanding of feature extraction.
The fine-tuned model is evaluated on a separate test set, achieving an accuracy of 79.57%. A baseline model is also evaluated, providing a comparison metric for the effectiveness of the fine-tuning process.
This README provides an overview of the codebase, its functionalities, and the results obtained. For detailed implementation and usage instructions, please refer to the corresponding code files.