Using Context Encoders in AEC/FM

Training Convolutional Neural Networks (CNNs) to predict the missing region in an image through the surrounding information is a very recent methodology in computer vision. Context Encoders is one of the state-of-the-art... [ view full abstract ]

Training Convolutional Neural Networks (CNNs) to predict the missing region in an image through the surrounding information is a very recent methodology in computer vision. Context Encoders is one of the state-of-the-art unsupervised learning algorithms that is able to perceive semantic of an image and generate the latent pieces. We have discovered that Context Encoders requires a great amount of training images to yield decent results. However, there are times when the application of Context Encoders is to be used on a new subject area without an adequate amount of training images. The goal of this work is to apply an adapted architecture and a convolutional neural network structure to successfully eliminate the noise from a particular object in images and inpaint the missing. To accomplish the above objective, we investigated an efficient unsupervised visual learning algorithm. Instead of exerting the encoder-decoder model, we employed the U-Net architecture, which can predict precise segmentation results from relatively few training images. Additionally, a derived convolutional neural network structure from the structure originally applied in U-Net is used in this algorithm. Although the algorithm we present is trained from relatively few images, it can efficiently generate a delicated result of the missing parts in an image. This research focus on erasing people from a particular building and produce new pixels from surround clues with encouraging results.