New carpet pattern design with deep learning

Error rates of the generator and discriminator network in the model coded with the neutral network AI application in carpets design

The fact that digitalization touches every aspect of today’s world has often been realized with the ideas of the industry and projects carried out for the development of the industry. While Artificial Intelligence shows its effective activities in many areas of the industry and it can also support designers in pre-production design. Within the scope of this study, it is aimed to produce new product designs by using the image data of the most demanded products obtained from a carpet manufacturing company with Deep Convolutional Generative Adversarial Networks. Before performing the model training with the images obtained from the carpet manufacturer company, the image generation performances of a ready-made data set and two different Python libraries, Keras and PyTorch, were compared. As a result, it was determined that the performance of the PyTorch Python library in generating images was higher quality, and the model built with real images was repeated. The synthetic designs produced were presented to the carpet manufacturer, and it was aimed to bring the role of the designer to a position that guides the design models with the designer’s experience and knowledge in the design process, which is a complex and stochastic process before production.

 

While discussing with designers and manufacturers how many projection designs to produce for only one season, approximately five-hundred projections are designed by about twenty designers is learnt. Nonetheless, all designs could not be made actual, which means that a huge workload on designers. Starting from this idea, we planned to benefit from Generative Adversarial Network, which is nowadays used in various areas such as producing images or detecting fraud in an image, for the design of a commercial product, a carpet.  Though there are different Python libraries, in this study before deciding to use which libraries for the model, we chose to compare two of them, which are Keras and PyTorch. After determining PyTorch is the most effective library for our study, conducting a serious of training with the sample dataset. After that, synthetic images produced in different libraries were compared over production speed, error rate and performance metrics in a graph. In conclusion, Deep Convolutional Generative Adversarial Network was coded with PyTorch on a real dataset, given the carpet manufacturer company, for the study and 64 synthetic images are produced. Then generated images were served to the company.

 

Figure 4 Sample images from the “African Fabric Images” sample data set used in the training

the sample data set "African Fabric Images", which is a free data set provided on the Kaggle platform, was used before synthetic designs were produced with the real data set. There are 1059 3D colored carpet/fabric images in the sample data set. In the created Generative Adversarial Network model, all 1059 data in the data set were used during the training. 64 images from the images in the data set are included in Figure 4 as an example.

Full article: New carpet pattern design with deep learning, August 2022, DOI:10.36909/jer.16781