Utilizing  pre-trained networks
for tasks of architectural image
generation and recognition. 

Advisor/   Sean Hanna and Kahlid El-Ashry



Image 01 -  A method of viewing how the network is understanding the similarity of the images after being trained. Architectural imagery on the right and architectural drawings on the left. Interestingly sections tend to be in the middle of this diagram.


Training deep neural networks to distinguish images is a well-documented practice, such as telling the difference between different types of dogs or types of flowers. This task is accomplished through feeding the network a surplus of well-structured labeled image data. However, for functions with sparse and difficult to label data, parallel research suggests, pre-trained networks can transfer their knowledge and accomplish a similar domain task with less training time and data. Often this method yields more accurate validation and accuracy scores overall.

Unfortunately, for issues grammatical in nature, such as natural language processing or architectural spatial understanding, the task presents a further challenge to these networks. Jeffery Elman, a cognitive scientist from UCSD, referred to this as the “projection problem” - That the example data might not be enough information for the network to generalize and understand the underlying logic of the data.

The attempt of this dissertation evaluates to what extent can pre-trained networks help reduce the information needed to distinguish architectural plans from sections and understand spatial patterns? Furthermore, in the process of domain adaptation, we uncover to what extent and wherein the layers of the network, the adjustment occurs from source to target domains. The method devloped to Fine-Tuning the networks with varied features spaces from four auxiliary datasets, flowers, art drawings, sketches of animals, and images of pets.

The documentation of the networks learning architectural spatial patterns in drawings yields surprising results. For specfic types of auxiliary data sets that most pertain to organization and to what distance the layers of the network adjust, please contact for results.

Additionally, we introduce a method to use generative adversarial networks GANs to create architectural drawings using pre-trained networks. Please contact for specific content and result information.

This work is on going













Image 02 – ALV Kernel _ An Image of the first layer inside a Network, depicting how the network is filtering and breaking down the image for a recognition task. We can see that at this stage the image is being understood for its contours. The color tone variance refers to what the excitatory in white and inhibitory in black. Input image of flower and dog





Image 03 – A comparison of network layers pre-trained depicting how the layers get into finer and finer details and features the deeper into the network. The layers are more similar due to the pre-trained developing general biases. Input image sunflower, dog, sketch of cat and architectural CAD plan.


Image 04 – A comparison of network layers trained from scratch depicting how the layers get into finer and finer details and features the deeper into the network. Input image sunflower, dog, sketch of cat and architectural CAD plan.