Cumincad : CUMINCAD Papers : Search Results

Hits 1 to 1 of 1

Reformat results as: short short into frame detailed detailed into frame

_id	caadria2023_106
id	caadria2023_106
authors	Li, Yuqian and Xu, Weiguo
year	2023
title	Research on Architectural Sketch to Scheme Image Based on Context Encoder
doi	https://doi.org/10.52842/conf.caadria.2023.1.069
source	Immanuel Koh, Dagmar Reinhardt, Mohammed Makki, Mona Khakhar, Nic Bao (eds.), HUMAN-CENTRIC - Proceedings of the 28th CAADRIA Conference, Ahmedabad, 18-24 March 2023, pp. 69–78
summary	Architects are used to hand drawing sketches to express the architectural creation intention. To present these abstract sketches, architects and teams need to convert sketches into architectural scheme images, which requires a lot of time and labour. Deep learning may have the potential to improve the efficiency of this work. The common sketch-to-image generation is based on Generative Adversarial Network (GAN), and the research of edge-to-image has made a big progress. But these methods require strict alignment of data pairs, which is difficult to achieve. Zhu et al. proposed the loss of Cycle-Consistent, which solved the problem that pairs of data sets are difficult to collect. However, most of the image translation methods require strict alignment between image data pairs, which can be achieved only for the edge mapping extracted from the image; but the sketch is very different from the edge. Due to the abstractness and fuzziness of the sketch, any simple distortion cannot complete the task of providing pixel-level alignment between the sketch and image; And image translation is the transfer of image features such as colour and texture. The original image has a strong constraint on the generated image, which makes the original structure of the image impossible to be changed. By image inpainting, we address this topic using a joint image completion approach with Context-Encoder, where the architectural sketch provides the image context for generating the scheme images. This setting has two advantages: first, the joint images can avoid the complexity of cross modal problems and the strict alignment of the data pairs as image-to-image translation; second, because of the weak constraint, the outputs have greater freedom, which perhaps can generate more imaginative results. The Context-Encoder generates scheme images on the data sets of general architectural sketches. The results present that the applicability of the completion method is better than that of the method of image translation. And scheme images that is different from the original architectural sketch contours have been generated.
keywords	Sketch, Building Scheme Image, Image Completion, Context-Encoder
series	CAADRIA
email
last changed	2023/06/15 23:14

No more hits.