Tehnički vjesnik, Vol. 31 No. 6, 2024.
Izvorni znanstveni članak
https://doi.org/10.17559/TV-20240416001467
Improved Discriminator Network Based on Residual Network for Person Image Synthesis
Zixuan Chen
; School of Computer Science, Hu Bei University of Technology, Wuhan, 430086, China
Lingyu Yan
; School of Computer Science, Hu Bei University of Technology, Wuhan, 430086, China
*
Chunzhi Wang
; School of Computer Science, Hu Bei University of Technology, Wuhan, 430086, China
Zhiwei Ye
; School of Computer Science, Hu Bei University of Technology, Wuhan, 430086, China
* Dopisni autor.
Sažetak
As the internet industry continues to advance, there has been a significant growth in the volume and variety of data, rich in features and diversity. Analyzing the distribution pattern of images quickly has become a challenge. The utilization of Convolutional Neural Networks (CNNs) has greatly enhanced the efficiency of data analysis and processing in areas like image recognition, image segmentation, and image synthesis. Despite their effectiveness, CNNs face challenges in terms of computational resources, convergence speed, and generating high-quality synthetic images. To address the described problems, this paper focuses on deep residual networks and generative adversarial networks in image synthesis algorithms. It divides human object image synthesis into three stages. The first step segments the human object image's target subject, the second step enhances the synthesized image data, and the third step fuses the segmented feature maps. To mitigate the issue of poor quality composite images, the DENSE-GAN method can be employed. To tackle the noted shortcomings, one approach is to enhance the discriminator network using a residual network. This involves substituting the CNN within the discriminator network's feature extraction module with a residual network. Additionally, to further enhance feature expansion, it is crucial to increase the network's depth by adding more layers. Different parts of the network output layer extract the overall contour information, local detail information, and identity information of the composite image. The results of these three modules are used to guide the weights of DENSE-GAN, adjusting them and experimenting on the dataset, which significantly improves the quality and clarity of DENSE-GAN composite images. Our algorithm achieved promising results on the MNIST, Market-1501, CelebA, and DeepFashion datasets. For example, after training for 100 epochs on the MNIST dataset, our algorithm exhibited an average loss of approximately 0.75.
Ključne riječi
DENSE-GAN; generative adversarial networks; image synthesis; residual networks
Hrčak ID:
321902
URI
Datum izdavanja:
31.10.2024.
Posjeta: 6 *