文献类型：期刊文献

英文题名：EnrichGAN: Exploiting enriched discriminator representations for training GANs under limited data

作者：Mu, Wenhao[1];Chen, Kai[2];Wang, Yihong[3];Ma, Lizhuang[4];Wang, Nan[1];Jiang, Qingchao[1];Huang, Bingcang[2]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China;[2]Gongli Hosp Shanghai Pudong New Area, Dept Radiol, Shanghai 200135, Peoples R China;[3]East China Univ Sci & Technol, Sch Math, Shanghai 200237, Peoples R China;[4]Shanghai Jiao Tong Univ, Dept Comp Sci & Engn, Shanghai 200240, Peoples R China

年份：2026

卷号：699

外文期刊名：NEUROCOMPUTING

收录：;EI(收录号：20262721049233);Scopus(收录号：2-s2.0-105043678890);WOS:【SCI-EXPANDED(收录号:WOS:001819645400001)】；

基金：This work was supported by the National Natural Science Foundation of China [grant numbers 62402181, 62322309] ; the Shanghai Magnolia Talent Plan Pujiang Project [grant number 24PJD022] ; and the Science and Technology Innovation Plan of Shanghai Science and Technology Commission [grant number 23S41900500] .

语种：英文

外文关键词：Generative adversarial networks; Scale-variant attention; Instance-level reconstruction; Mode-sensitive differential

摘要：Generative adversarial networks (GANs) have demonstrated significant efficacy in high-quality visual synthesis. However, model performance remains dependent on the availability of large-scale training datasets. Under limited data conditions, the discriminator is highly susceptible to overfitting, which degrades its capacity to learn effective feature representations and provide informative gradient guidance for the generator. This deficiency prevents the comprehensive capture of the underlying data distribution and frequently induces mode collapse. To address this representational bottleneck, this paper proposes a unified framework named EnrichGAN, which enhances synthesis performance by enriching the intrinsic representation of the discriminator. Specifically, a scale-variant attention module is introduced to refine the extraction of critical semantic features, thereby mitigating overfitting to local discriminative information. Subsequently, the discriminator reconstructs each image instance from these refined features to establish a shared feature manifold between the real and fake distributions. Utilizing this learned shared feature manifold, a mode-sensitive differential is applied to both the discriminator and generator to encourage comprehensive coverage of the underlying real distribution and facilitate the discovery of minor modes. Extensive experiments demonstrate that the proposed EnrichGAN framework consistently outperforms state-of-the-art methods, validating its effectiveness in achieving high-quality and diverse image synthesis under limited data scenarios.