Zhimeng He

Through the Eyes of Involution

This image was generated during the development of my Trans-Involution deep learning network for building rooftop extraction. The model integrates Involution blocks, which guide the network to focus on the most informative regions of the image—such as structural edges and meaningful textures—enhancing its ability to distinguish buildings from their surroundings.

These two images represent different channel responses within the Involution block of the network. The left image highlights how certain channels focus specifically on building structures, capturing edges and roof textures. In contrast, the right image shows channels that emphasize vegetation, such as grass and trees, helping the model distinguish between natural and man-made elements in complex urban environments.

This visual not only reflects how the model "sees" urban structures but also provides a glimpse into the interpretability of AI vision. Through this piece, I aim to bridge the gap between machine perception and human aesthetics, showcasing the inner workings of deep learning in a visually compelling way.

I hope this artwork helps humans better understand what AI "sees," and sparks curiosity about the beauty that can emerge when technology and creativity intersect.Text content