Unsupervised geometric classifier based on optical feature extraction

Diffractive deep neural networks (D2NNs) offer extraordinary advantages for applications in real-time object classification, computational imaging, and optical computing, where speed and energy efficiency are critical. Unsupervised learning reduces reliance on labeled data, enhances robustness, and is inherently compatible with optical parallel processing. In this work, an unsupervised geometric classifier combining optical feature extraction and clustering is proposed. A D2NN processor trained on the Laguerre-Gaussian modes serves as an optical encoder, mapping the input fields to a six-dimensional feature space. Circles, along with randomly rotated triangles, squares, and pentagons, are modulated onto the phase of a Gaussian beam. After the modulated beams pass through the D2NN-based encoder, the resulting feature vectors enable unsupervised classification via clustering. To the best of our knowledge, this is the first demonstration of unsupervised classification of different geometric shapes based on optical feature extraction. The experimental results show 100% classification accuracy for the four different geometric shapes, including circles and randomly rotated triangles, squares, and pentagons.