Implementation of GANs as a tool in Fashion Industry

Project Title

Implementation of GANs as a tool in Fashion Industry

Project Area of Specialization Artificial IntelligenceProject Summary

In this era of multimedia communications and its uses in all aspects of daily life, it is important to assist the designers in creating innovative and useful designs for various applications. In this work, we develop a system to generate new patterns and designs that are based on artificial intelligence paradigm, which assist professionals in creating images suitable for a particular design applications. The proposed system can also be used to identify existing images for their authenticity and classify them as real of fake.

The computer vision techniques are applied to process the images in combination with deep learning approach, such as Generative Adversarial Networks (GANs). Recently, GANSs have shown desirable properties and suitability for learning applications, in particular, for artificial intelligence applications involving image processing. GAN’s basically consist of two parts:

1. Generative: This part employs image processing methods in conjunction with deep learning algorithms to generate new images with desired properties. The designer assists the algorithm in learning by providing a set of suitable parameters and seed images. As a result, the algorithm performs extensive and rigorous calculations to implement learning algorithms to create new images. One of the applications of this algorithm is to generate images, designs and patterns for textile industry.
2. Discriminative: In this part, the proposed image processing and learning system processes the existing images to identify the authenticity by classifying them as fake or real. As an example, a fake currency note can be identified using the proposed algorithm.

In creating new patterns and designs, there is a chance of plagiarism as the seed images are used from the existing library. The discriminative part ensures that the results of new images by the generator are different and unique. We further enhance the proposed system in learning and performance by introducing convolutional layers in the generator network called Deep Convolution GAN’s (DCGAN).

Project Objectives

Textile industry constitutes around 54% of Pakistan's revenue which is extremely important for the country's economy. This project will enable textile industry to benefit form new designs, which will contribute to the economy. This project can assist them in generating new patterns without that much of an effort. It is also unsupervised so it doesn’t need human attention all the time. The time and man power saved can further be used in any other area.

The system is applied to the drug design. This will enable the designers to explore more possibilities of drug design. 

The proposed system is applied to the defense industry.

Project Implementation Method

An important aspect in the use and implementation of the proposed system deals with the amount of assistance/supervision provided to the learning algorithms. This varies from absolutely no assistance in learning to a considerable contribution by humans. The implementation is divided into two categories:

Supervised Learning:  Supervised learning is used only for labeled/organized data. Examples are classification and regression.

Unsupervised Learning: The unsupervised learning is widely applied where data is unlabeled or is difficult to label. Examples are clustering, data compression etc. We are using unsupervised learning for two reasons, one reason is that it is not prohibited to think of those new cases that can happen while solving the problem, the other and the more specific one is that our data is critically unlabeled and will take a lifetime to label manually.

The implementation of the proposed system consists of the following steps:

Data Collection: Data collection is the most fundamental part of this project. In order to acquire image data from various sources, it is important to consider copyright conditions and permissions associated with the images for commercial/research purposes. The performance of the algorithm depends on the input seed images and the characteristics exhibited by them. A considerable amount of diversity is required in the images that are required to be used for the learning. The most common source of input images are the various online databases that provide free images that may be used to test and train the proposed algorithm. In addition to the availability, the quality of images plays a vital role in determining the quality of learning and the resultant generated images. It is important to use high resolution and clear images, which is sometimes difficult to acquire in large numbers. The patterns in the seed images need to be random and distinct.

Implementation: In the implementation process, the dataset is processed by the discriminative part of the algorithm while noise is treated by the generative portion of the algorithm. The output from the generative part is fed into the discriminative area where the distance between the distributions of original and the generated image is calculated. Based on the distance between two distributions, the images are classified as 0 (for fake) and 1 (for original). The decision is revised and weights are updated iteratively, which is based on the results of the back propagation.

In the beginning, the results are arbitrary and noisy and there is a huge difference between the distributions of the real and generated image, but as the training proceeds the loss mitigates that improve and images. As the images evolve, the new versions start to show features. The training is carried on until the distributions are close enough and loss is negligible.

RESULTS in the start:

 RESULTS after some training:

___________________

Benefits of the Project

In the first phase of the implementation, the main objective is to cater the textile industry of Pakistan by designing new patterns and textures. In order to use the new patterns, it is ensured that the designs do no resemble any existing work. In particular, the implementation of the algorithm is applied to designing dresses, style transfer, image retrieval, bedrooms designing, super resolution, and face frontal view generation.

In the second phase, this deep learning will be adapted for drug design in the molecular biology area. It is expected that the drug design process will be accelerated by employing deep learning techniques. The impact on the quality of human lives is expected to be high.

The proposed system can also be beneficial in improving our defense system, designing new components, infrastructure and many other advancements can be made using GAN’s.

Technical Details of Final Deliverable

The Graphical Processing Units (GPUs) are required to run computationally complex algorithms in an efficient manner. The entire process needs a very good Graphic Processing Unit that can manipulate matrices and graphics quickly and efficiently. The following hardware/chipset is required to design and implement the proposed system.

VIDIA GTX 1080:
Memory Size: 8 GB,         
Memory bus: 256 bit
Memory bandwidth: 320 GBs/second
Processing power: 2560 cores,
GPU clock:  1607MHz

The final deliverable includes a system that uses GPUs to efficiently generate new patterns and designs to be used in textile industry. In the second phase, the system will be adapted to assist in drug design and analysis. The system is further adapted for applications in defense industry.

Final Deliverable of the Project HW/SW integrated systemType of Industry Manufacturing , Security Technologies Artificial Intelligence(AI)Sustainable Development Goals Industry, Innovation and Infrastructure, Peace and Justice Strong InstitutionsRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	80000
NVIDIA GTX 1080: Memory Size: 8 GB	Equipment	1	70000	70000
Printing supplies, Stationary	Miscellaneous	2	5000	10000