Structural edge detection of photographic images of rooms with machine learning

In the task to reconstruct 3D models of room architecture from photographic images, identifying the relevant structural edges of the room amidst the noise has been a tremendous challenge. This Final Year Project sets out to determine if machine learning can be a viable alternative to classical edge-detection algorithms and, if so, determine the machine learning model that has the best performance. The methodology for this project involves four main parts – generating a labelled dataset, augmenting the labelled data to enlarge the dataset, processing the dataset, and training various models on the dataset. For this project, training is carried out on four different Fully Convolutional Network (FCN) architectures, namely SegNet, U-Net, DenseNet and a pre-trained FCN-RESNET101. For each model, the input is an RGB image and the output is a greyscale image with each pixel indicating its probability of not laying on a relevant edge. The results obtained from the training indicated that the model based on U-Net had the best performance out of the four. Using this finding, further finetuning of the U-Net model’s parameters and hyperparameters are performed to further enhance its performance. Post-processing such as edge-thinning and feature extraction is applied to the output of the final model to obtain the line equation of every predicted edge. The results obtained showed strong promise in discerning structural edges, thus validating the initial hypothesis that machine learning is a viable alternative to classical algorithms. Future works include further enhancing the current model’s accuracy and creating an algorithm to construct a wireframe model from the line equations.Bachelor of Engineering (Mechanical Engineering