This Multi-Class Classification network, implemented with Keras, is used to develop a Computer Vision model capable to classify architectural drawings.
The aim of this project is to make the power of Computer Vision useful to automatically sort and analyze the building-related data. Here, a CNN is trained to distinguish between three types of architectural drawings:
elevationfloor-plansection
This project is part of an ongoing research from the FID BAUdigital, conducted at the Universitäts- und Landesbibliothek Darmstadt, from the TU Darmstadt.
Due to copyright issues, the data used to train the models can not be publicly shared in this repository.
multi_class_classification.py
Data preparation was done through the flow_from_directory method. All images have to sorted into different folders according to the categories. The CNN uses softmax activation in the last layer and 3 neurons as output to enable the multi class classification. Useful Callbacks are installed, like tensorboard or Learning Rate Scheduler. Don't forget to adjust the learning rate in the lr_time_based_scheduler-function. In this file a time-based decay is used. Please play with the model's architecture, change the amount of layers or width, to optimize performance.
predict_classes.py
After successful training of the model, use this file to run predictions on either a single image sample or multiple images (both functions available). This creates the drawings relevant metadata in a machine-readable format The predictions will be converted to a list containing all information, using the pandas DataFrame method. Available formats are: .csv / .xlxs / .json
continue_training.py
This file can be used to optimize the trained model. You can use different data or model architecture to optimize the layer weights.
For questions or further ideas, email me at paul.arch@web.de