CV-Project

Warning: Some content is NSFW (Not Safe For Work).

The dataset folders "originals" and "photoshops" contain material that may be offensive or inappropriate for some audiences. Viewer discretion is advised.

Performance Metrics

Model	1000 Train	1000 Test	1000 F1	5000 Train	5000 Test	5000 F1	10000 Train	10000 Test	10000 F1
ResNet50	0.88	0.64	0.61	0.88	0.72	0.71	0.93	0.88	0.87
VisionTransformer	0.91	0.68	0.60	0.92	0.81	0.79	0.94	0.90	0.87
InceptionV3	0.90	0.71	0.65	0.88	0.75	0.67	0.92	0.89	0.88
GoogLeNet	0.91	0.68	0.68	0.77	0.71	0.72	0.85	0.83	0.82
DenseNet	0.98	0.67	0.64	0.93	0.72	0.70	0.89	0.84	0.82
ResNeXt50	0.99	0.62	0.61	0.97	0.76	0.72	0.94	0.89	0.88
ELA + ResNeXT (Ours)	0.92	0.75	0.69	0.93	0.83	0.82	0.98	0.94	0.93

Results for the models: ResNet50, Vision Transformer, InceptionV3, GoogLeNet, DenseNet, ResNeXt50. Run all the blocks of "classifier.ipynb" notebook for train and test accuracy.

Results for the model "ELA + ResNeXT": Run the "data.ipynb" notebook first to preprocess the raw RGB images to generate corresponding error level analysis of the image. Then run all the blocks of "ELA_ResNeXt_Classifier.ipynb" for train and test accuracy.

Run the "plots.ipynb" notebook for Train Accuracy plot and confusion matrix of the best performing model.

Project Poster: