Performance Evaluation of Generative Adversarial Networks for Computer Vision Applications

• Published in: Ingénierie des systèmes d'Information Vol. 25; no. 1; pp. 83 - 92
• Main Authors: Nannapaneni, Sravani, Sistla, Venkatramaphanikumar, Kolli, Venkata Krishna Kishore
• Format: Journal Article
• Language: English
• Published: Edmonton International Information and Engineering Technology Association (IIETA) 01.02.2020
• Subjects: Artificial neural networks; Belief networks; Competition; Computer vision; Datasets; Discriminators; Distribution functions; Generative adversarial networks; Handwriting; Image classification; Image processing; Image segmentation; Neural networks; Optimization; Performance evaluation; Probability; Probability distribution; Probability distribution functions; Random variables; Task complexity
• ISSN: 1633-1311; 2116-7125; 2116-7125
• DOI: 10.18280/isi.250111

Generative Adversarial Networks (GAN) generates model approaches using Convolution Neural Networks (CNN) to find out learning regularities and to discover the hidden patterns held in given input data. GAN is a generative model that is trained using two models such as generator and Discriminator both competing against each other to learn the probability distribution function, networks such as CNN, RNN, ANN etc. These traditional neural networks are easily fooled in misclassifying things by adding small amount of noise to original data, whereas GAN’s are more stable and easier to train due to the amalgamation of Feed Forward Neural Network and CNN. In general, GAN’s are simple Neural networks be trained in adversarial way to generate the data mimicking same distribution, Generator learns new possible sample, and the Discriminator learns how to differentiate generated samples from valid facts. Generated samples are similar in the nature but different from real distribution data. The generated samples make use of computer vision techniques such as visualization designs, realistic image generation, image classifications etc. In the proposed work, to realize the probability distribution Restricted - Boltzmann machines and Deep Belief networks are used. The performance of the GAN Networks is evaluated on various standard datasets to realize the complex tasks such as image prediction, handwritten digit’s generation, clothing classification, image segmentation tasks etc. From the experimental results, it is clearly evident that the performance of GAN outperforms other state of the art classifiers on all the benchmark datasets.