Fire-Image-Classification

This project addresses a predictive problem of classifying images as fire or non-fire to aid fire departments. The features are image pixel values, with a binary fire/non-fire label as the target variable. We will compare three different machine learning techniques: CNN, SVM and KNN. CNNs have layers which are useful in predicting patterns. K-nn follows an easy implementation and is very robust to noise which is a very common byproduct of these datasets. Considering the complexity of an image base data, SVM’s are a great model as well. Since, SVM and KNN are not able to handle high dimensional data quite well, they will be trained on PCA-reduced data. Preprocessing will involve image resizing, normalization, and PCA for dimension reduction. We will implement class balancing to ensure a balanced distribution between fire and non-fire images using techniques like oversampling or undersampling(less likely). We will use cross-validation, graphical analysis for evaluating our model to ensure robust model performance. The evaluation strategy will help in identifying if our model is overfitting or underfitting. The focus on recall ensures that the model minimizes false negatives, which is crucial in fire detection to avoid missing actual fire incidents. There are a few existing solutions available but those solutions use pre-trained weights and don’t compare different classifications techniques. Libraries we plan to use for this project: Pytorch, numpy, sklearn, pandas, matplotlib, seaborn, pickle, PIL.