Fine-tuned DenseNet for Image Classification

Model Overview

This fine-tuned DenseNet121 model is designed to classify images into the following categories:

DALL-E Generated Images
Human-Created Images
Other AI-Generated Images

The model is ideal for detecting AI-generated content, particularly useful in creative fields such as art and design.

Use Cases

AI Art Detection: Identifies whether an image was generated by AI or created by a human.
Content Moderation: Useful in media, art, and design industries where distinguishing AI-generated content is essential.
Educational Purposes: Useful for exploring the differences between AI and human-generated content.

Model Performance

Accuracy: 95% on the validation dataset.
Loss: 0.0552 after 15 epochs of training.

Training Details

Base Model: DenseNet121, pretrained on ImageNet.
Optimizer: Adam with a learning rate of 0.0001.
Loss Function: Cross-Entropy Loss.
Batch Size: 32
Epochs: 15

The model was fine-tuned using data augmentation techniques like random flips, rotations, and color jittering to improve robustness.

Training Metrics

1. Loss Over Epochs

This graph shows the decrease in loss over 15 epochs, indicating the model's improved ability to fit the data.

2. Accuracy Over Epochs

This graph shows the increase in accuracy, reflecting the model's growing ability to correctly classify images.

Sample Dataset

Here is a visual representation of the dataset used for training and validation:

This image shows a collage of examples from the dataset used to fine-tune the DenseNet model. The dataset includes a diverse mix of images from three distinct categories:

Human-Created Images – Traditional artwork or photographs made by humans.
DALL-E Generated Images – Images created using DALL-E, an advanced AI model designed to generate visual content.
Other AI-Generated Images – Visual content generated by other AI systems, aside from DALL-E, to provide variety in the training data.

This diversity allows the model to effectively learn how to distinguish between different forms of image creation, ensuring robust performance across a range of AI-generated and human-created content.

Model Output Samples

Here are some examples of the model's predictions on various images:

Sample 1: Human-Created Image

Predicted: Human-Created

Sample 2: DALL-E Generated Image

Predicted: DALL-E Generated

Sample 3: Other AI-Generated Image

Predicted: Other AI-Generated

Model Architecture

Feature Extractor: DenseNet121 with frozen layers to retain general features from ImageNet.
Classifier: A fully connected layer with 3 output nodes, one for each class (DALL-E, Human-Created, Other AI).

Limitations

Data Bias: The model's performance is dependent on the balance and diversity of the training dataset.
Generalization: Further testing on more diverse datasets is recommended to validate the model’s performance across different domains and types of images.

Model Download

You can download the fine-tuned DenseNet121 model using the following link:
Download the Model

References

For more information on DenseNet, refer to the original research paper:
Densely Connected Convolutional Networks (DenseNet)