GAN-BASED PERSON REIDENTIFICATION USING DWT

Person Re-identification (Re-id) is used in real-world applications for non-invasive biometric surveillance, validation, and identification of people in congested areas. In this paper, we propose GAN-based Person Reidentification using DWT. Generative Adversarial Network (GAN) generates identical persons with altered postures that appear in the output images. For every picture of a person, eight new photos are produced by defining a set of eight predefined poses. In order to extract features, the Discrete Wavelet Transform (DWT) technique is used which breaks down images into three high-frequency bands and one low-frequency band. The one low-frequency band of one-fourth of the size of the original image is considered for the final features with less noise. The features of test and database images are compared using Cityblock, Euclidean Distance (ED), and Cosine Similarity approaches. Through enhanced feature discrimination made possible by this dual technique, people are effectively re-identified from various camera angles and improve recognition accuracy by up to 99%.