AlexNet_IN1k
AlexNet turns images into numerical representations that can be compared for visual similarity. In EIDORA, it is a useful starting point for exploring image collections or comparing a well-established image model with newer approaches.
Intended Use
- Explore visual similarities: find images with related visual patterns in a collection of photographs or artworks.
- Compare methods: use AlexNet as a baseline alongside other image embedding models.
AlexNet works with images only. If you want to search images using written descriptions, a text–image model such as CLIP is a better fit.
Research Example
Brachmann, Barth and Redies (2017) used features from an AlexNet-based network to study visual properties of Western, Islamic and Chinese artworks. They examined whether patterns in the network's responses could distinguish artworks from other kinds of images. Their study used a modified network and different features, not the exact EIDORA embedding. Read the paper.
Limitations
AlexNet learned its features from photographs labelled for object recognition, so similarities between its outputs do not necessarily match art-historical categories or cultural meanings. The centre crop may also remove details near image edges. Inspect the retrieved images before interpreting the results; similarity alone does not establish authorship, influence or historical connection.
Technical Details
- Architecture: convolutional neural network (CNN); originally introduced by Krizhevsky, Sutskever and Hinton (2012).
- Checkpoint used: TorchVision
AlexNet_Weights.IMAGENET1K_V1, pretrained on ImageNet-1K. This is TorchVision's implementation and checkpoint, not the original authors' 2012 weight file. - Input: RGB image; resize the shorter side to 256 pixels, then take a 224 × 224 centre crop (bilinear interpolation). ImageNet channel normalization is included in the ONNX graph.
- Output: one L2-normalized, 4,096-dimensional embedding per image, taken from the second fully connected layer after ReLU. Compare embeddings with cosine similarity.
- Model size: approximately 228 MB for this ONNX conversion.
- Attribution: EIDORA provides the ONNX conversion of TorchVision's pretrained model. TorchVision's code is BSD-3-Clause licensed; the pretrained weights' redistribution terms should be considered separately.
References
- Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems, 25.
- Brachmann, A., Barth, E., & Redies, C. (2017). Using CNN Features to Better Understand What Makes Visual Artworks Special. Frontiers in Psychology, 8, 830.
- Russakovsky, O., et al. (2015). ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115, 211–252.
Citation (BibTeX)
@inproceedings{krizhevsky2012imagenet,
title = {ImageNet Classification with Deep Convolutional Neural Networks},
author = {Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E.},
booktitle = {Advances in Neural Information Processing Systems},
volume = {25},
year = {2012},
url = {https://proceedings.neurips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html}
}
@article{brachmann2017using,
title = {Using CNN Features to Better Understand What Makes Visual Artworks Special},
author = {Brachmann, Anselm and Barth, Erhardt and Redies, Christoph},
journal = {Frontiers in Psychology},
volume = {8},
pages = {830},
year = {2017},
doi = {10.3389/fpsyg.2017.00830}
}
@article{russakovsky2015imagenet,
title = {ImageNet Large Scale Visual Recognition Challenge},
author = {Russakovsky, Olga and Deng, Jia and Su, Hao and Krause, Jonathan and Satheesh, Sanjeev and Ma, Sean and Huang, Zhiheng and Karpathy, Andrej and Khosla, Aditya and Bernstein, Michael and Berg, Alexander C. and Fei-Fei, Li},
journal = {International Journal of Computer Vision},
volume = {115},
pages = {211--252},
year = {2015},
doi = {10.1007/s11263-015-0816-y}
}