DataScienceProject/Resnet

Real art vs AI-Generated art image classification

This project provides a ResNet50 pre-trained model for classifying images as either 'real art' or 'fake art'. ResNet50 is a deep convolutional neural network with 50 layers, known for its "residual connections" that help mitigate the vanishing gradient problem. It allows training of very deep networks by adding shortcut connections that skip one or more layers, making it highly effective for image classification tasks. Our goal is to accurately classify the source of the image with at least 85% accuracy and achieve at least 80% in the Recall test.

Installation instructions

The following libraries or packages are required: numpy, pandas, tensorflow, keras, matplotlib, sklearn, cv2. We prepare the data for the model by sorted the images into 2 types of folders which are divided equally(real art- labeled as 0, fake art- labeled as 1). Our ResNet50 model is based on 2,800 images that have been resized and normalized, the files formats is PNG‬, JPG‬. The images are divided into a training set that contains 90% from data and a testing set that contains the remaining 10%.

ResNet50 model architecture

The model is pre-trained on 'ImageNet' that contains a large dataset of more than millions images. It applies transfer learning, freezing initial layers of ResNet50, and training only the final layers. The final layer, which makes the predictions, is a binary classification layer that uses a sigmoid activation function.

Training Details

The model is trained using binary cross-entropy loss and the Adam optimizer. The model validates itself during training using 20% of the training data as validation, independent of the test data, to monitor performance and avoid overfitting. The model is trained for 5 epochs with a batch size of 32 and employs 4-fold cross-validation to ensure robust performance. During each fold, the model's weights are saved after training, allowing for the reuse of the best-performing weights.

Performance Evaluation

After training, the model is evaluated on the test set. The following metrics are used to measure performance:
Accuracy: The percentage of correct classifications. Precision, Recall, F1-Score: For evaluating the model’s classification ability on both real art and AI-generated art images. Confusion Matrix: Provides insights into classification performance. Displays true positives, false positives, true negatives, and false negatives.

To run the project

Place the images in the respective training and testing folders. Preprocess the images by resizing and normalizing them. Train the model using the provided code. Evaluate the model on the test set.

Visualization results

Confusion Matrix: To visualize the classification performance. Training and Validation Metrics: Plots for accuracy and loss over the epochs.

Results

Test accuracy = 0.784

Test loss = 0.48

Precision = 0.76

Recall = 0.83

F1 = 0.79

Confusion Matrix: