Image Similarity in Percentage %

Siamese network to compare image similarity in percentage - based on Keras deep learning model (VGG16, ResNet50) & cosine similarity, euclidean similarity

Accuracy

The cosine similarity and euclidean similarity are shown in the table.

image1	image2	cosine similarity (VGG16)	euclidean similarity (VGG16)	cosine similarity (ResNet50)	euclidean similarity (ResNet50)
		84.51%	0.01326	91.28%	0.05116
---	---	---	---	---	---
		63.95%	0.00980	54.98%	0.02871
---	---	---	---	---	---
		100.00%	1.0	100.00%	1.0
---	---	---	---	---	---
		63.66%	0.01222	78.96%	0.03771
---	---	---	---	---	---
		51.74%	0.01105	51.18%	0.02189
---	---	---	---	---	---
		23.80%	0.00907	30.91%	0.01755
---	---	---	---	---	---
		42.68%	0.01361	49.00%	0.02593
---	---	---	---	---	---
		69.20%	0.01478	70.07%	0.02849
---	---	---	---	---	---
		77.01%	0.02064	82.51%	0.04565
---	---	---	---	---	---
Original Image	Cropped Image	93.75%	0.03695	95.31%	0.07801
---	---	---	---	---	---
Original Image	Adversarial Image	74.47%	0.01384	90.14%	0.06188
---	---	---	---	---	---
Original Image	Adversarial Image	79.60%	0.01324	91.45%	0.04503
---	---	---	---	---	---
Original Image	Screenshot Image	97.95%	0.06415	98.69%	0.13120

License

• This image similarity model is developed based on Keras deep learning model from https://github.com/fchollet/deep-learning-models
• The ResNet50 weights are ported from the ones released by Kaiming He under the MIT license.
• The VGG16 and VGG19 weights are ported from the ones released by VGG at Oxford under the Creative Commons Attribution License.
• The Inception v3 weights are trained by ourselves and are released under the MIT license.
• Thanks to DAmageNet: A Universal Adversarial Dataset https://arxiv.org/abs/1912.07160

Citations

@inproceedings{Panagiotis2021,
  author = {Panagiotis Kasnesis, Ryan Heartfield, Xing Liang, Lazaros Toumanidis, Georgia Sakellari, Charalampos Patrikakis, George Loukas},
  booktitle = {Journal of Applied Soft Computing},
  title = {Transformer-based identification of stochastic information cascades in social networks using text and image similarity},
  year = {2021}
}