Özet:

Abstract Person Re-identification (Re-id) is helpful in society for non-invasive biometric person identification, validation, and surveillance, in crowded places. This paper proposes person reidentification based on Enhanced Position-Regularization Generative Adversarial Network (EPR-GAN) using GLCM, radon transform, and DCT. The EPR-GAN model produces output images of the same person with the new postures. A set of eight established poses are defined, and create eight new images for any given picture of a person. The texture analysis and a subspace learning approach are used for features using Gray-Level Co-occurrence Matrix (GLCM), radon transform, and Discrete Cosine Transform (DCT) on EPR-GAN generated images. The GLCM technique is adopted with matrix sizes of 4X4, 8X8, 16X16, 32X32, and 64X64 on the given resized image of 128X64 to discover the local details of the set of images. As the dimension of the GLCM increased, the rank-1 recognition also increased due to the large size; however, the GLCM matric dimension is limited to 64X64, i.e., 4096. The radon transform is the image intensity projection along a radial line oriented at a precise angle applied on the GLCM matrix to enhance the rank one recognition accuracy. The resultant feature values of radon transform applied on 64X64 GLCM matrix is 95X180, i.e., 17100 is very high. Furthermore, the dimensionality reduction algorithm DCT is used on radon transform to decrease the number of final compelling features and enhance the model’s performance. As a result, the proposed model’s efficiency is better than the existing models.

Anahtar Kelimeler: