Style-based generative adversarial networks for enhancing deep-learning-based person re-identification

Deep-learning (DL) technologies have greatly boosted the performance of person Re-ID. However, they add a new challenge, in that these deep methods need large amounts of labeled data for training. Hence, person Re-ID remains an open problem, for which no prominent solutions have been found for all the different challenges. Thus, there is a need to develop a person Re-ID method that uses DL technology with a sufficient dataset for training. This study proposes an improved person Re-ID method based on DL technology, along with a StyleGAN and LSRO algorithm. The proposed method can solve the main issue of DL-based person Re-ID, namely the lack of data needed for training. It begins by constructing a successful baseline model to extract the strong discriminative features necessary for person Re-ID by modifying a general CNN model developed for the general object recognition task. Then, fine-tuning it using the transfer learning approach to make it more suitable for the person Re-ID problem domain. Moreover, random erasing and re-ranking are combined with the baseline model proposed herein to achieve significant performance improvement further and avoid overfitting. Afterward, the proposed method for person Re-ID exploits the StyleGAN to generate synthetic images that are both new and high-quality using the person Re-ID datasets that already exist. These newly generated images are then used to enlarge the training sets by introducing more extensive variations in terms of illumination, background, poses, and color. Then, the LSRO algorithm is used to integrate the generated images into the originally labeled training images. This is done by giving each of them uniform label distribution and designating a regularized loss function to them for the training of the baseline model. Generation of the new high-quality synthetic images using the StyleGAN and integrating them with the real images in datasets using the LSRO is the foremost contribution of the present research. The conducted experimental analysis and results proved that the proposed person Re-ID approach yielded better overall performance when compared with state-of-the-art person Re-ID approaches. Keywords: Person re-identification; Deep learning; Transfer learning; Convolutional neural networks; Generative adversarial networks; Label smoothing regularization for outliers; StyleGAN.