Person Re-Identification Utilizing GLCM, Radon Transform, and LDA based on Generative Adversarial Network

Abstract

Person re-identification, commonly called Re-id, is an effective non-invasive biometric technique for identifying individuals, validating identities, and monitoring crowds globally.  This study proposes a method for human re-identification utilizing a combination of Gray-Level Co-occurrence Matrix (GLCM), Radon Transform, and Linear Discriminant Analysis (LDA) within a framework based on Generative Adversarial Networks (GAN).  The GAN model generates output images of the same individual in various new poses.  Each original image produces a series of eight predefined poses, resulting in eight unique photos.  Texture analysis and a subspace learning approach are utilized to extract features from GAN-generated images using GLCM, Radon Transform, and LDA.  For the scaled image size of 128x64, GLCM is employed with matrix dimensions of 4x4, 8x8, 16x16, 32x32, and 64x64 to identify local features within the image dataset.  As the dimensionality of the GLCM increased, rank-1 recognition improved significantly, although the GLCM matrix was limited to a dimension of 64x64, totaling 4096 elements.  To enhance the accuracy of rank-1 recognition, the GLCM matrix undergoes the Radon Transform, which projects image intensity along radial lines at specific angles.  The feature values generated from applying the Radon Transform to the 64x64 GLCM matrix result in a size of 95x180, yielding a total of 17,100, which is quite substantial.

Furthermore, incorporating LDA into the Radon Transform enhances the model's performance, leading to more robust features.  To demonstrate the effectiveness of the proposed approach, experiments were conducted and validated using two widely recognized datasets: the smaller CUHK01 and the larger Market-1501 public dataset.  The results indicate that the performance of the proposed model surpasses that of existing methods.