Neighboring Algorithm for Visual Semantic Analysis toward GAN-Generated Pictures.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) produce synthetic images, but evaluating their visual quality and distortions remains challenging.
• Existing GAN-generated image and character evaluation algorithms are limited in capability and efficiency.

Purpose of the Study:

• To propose a novel, efficient, and accurate algorithm for the automatic and extrinsic evaluation of GAN-generated images.
• To quantify the visual distortions in GAN-propagated portraits.

Main Methods:

• Utilized artificial neural networks to extract image boundaries, forming a homogeneous portrait candidate pool.
• Employed the K-nearest neighbors (KNN) algorithm to identify similar concepts within the candidate pool for quality scoring.
• Trained the generative similarity properties on diverse classical datasets.

Main Results:

• The proposed algorithm significantly enhances the efficiency and accuracy of evaluating GAN-generated images.
• Achieved a calculated metric 1/9-1/28 compared to existing methods.
• Increased consistency with human visual perception by over 80%.

Conclusions:

• The novel KNN-based algorithm provides an effective solution for objective visual quality assessment of GAN-generated images.
• The method demonstrates superior performance in terms of efficiency, accuracy, and human perceptual alignment.