《電子技術(shù)應(yīng)用》
您所在的位置:首頁(yè) > 通信與網(wǎng)絡(luò) > 設(shè)計(jì)應(yīng)用 > 基于生成對(duì)抗網(wǎng)絡(luò)的小樣本圖像數(shù)據(jù)增強(qiáng)技術(shù)
基于生成對(duì)抗網(wǎng)絡(luò)的小樣本圖像數(shù)據(jù)增強(qiáng)技術(shù)
網(wǎng)絡(luò)安全與數(shù)據(jù)治理 6期
楊鵬坤,,李金龍,郝潤(rùn)來(lái)
(中國(guó)科學(xué)技術(shù)大學(xué)計(jì)算機(jī)科學(xué)與技術(shù)學(xué)院,,安徽合肥230026)
摘要: 基于生成對(duì)抗網(wǎng)絡(luò)(GANs)的圖像數(shù)據(jù)增強(qiáng)方法在近年來(lái)展現(xiàn)出了巨大的潛力,。然而生成高分辨率、高保真圖像通常需要大量訓(xùn)練數(shù)據(jù),,這和缺乏訓(xùn)練數(shù)據(jù)的現(xiàn)狀背道而馳,。為解決這一問題,提出了一種能夠在小樣本,、高分辨率圖像數(shù)據(jù)集上穩(wěn)定訓(xùn)練的條件生成對(duì)抗網(wǎng)絡(luò)模型,并且將該模型用于數(shù)據(jù)增強(qiáng),。實(shí)驗(yàn)結(jié)果表明,在基準(zhǔn)數(shù)據(jù)集上,,該模型與當(dāng)前最新模型相比能夠生成更加逼真的圖像并取得了最低的FID值,;在圖像分類任務(wù)中使用其進(jìn)行數(shù)據(jù)增強(qiáng)能夠有效緩解分類器的過(guò)擬合問題。
中圖分類號(hào):TP391
文獻(xiàn)標(biāo)識(shí)碼:A
DOI:10.19358/j.issn.2097-1788.2023.06.013
引用格式:楊鵬坤,,李金龍,,郝潤(rùn)來(lái).基于生成對(duì)抗網(wǎng)絡(luò)的小樣本圖像數(shù)據(jù)增強(qiáng)技術(shù)[J].網(wǎng)絡(luò)安全與數(shù)據(jù)治理,2023,,42(6):79-84,,102.
Few-shot image data augmentation based on generative adversarial networks
Yang Pengkun,Li Jinlong,,Hao Runlai
(School of Computer Science and Technology, University of Science and Technology of China, Hefei 230026, China)
Abstract: In recent years, image data augmentation methods based on Generative Adversarial Networks (GANs) have shown great potential. However, generating highresolution, highfidelity images typically requires a large amount of training data, which contradicts the current lack of training data situation. To address this issue, a conditional GAN model that can stably train on fewshot, highresolution image datasets has been proposed for data augmentation. Experimental results on benchmark datasets indicate that this model, compared to the current stateoftheart models, is capable of generating more realistic images and achieving the lowest Fréchet Inception Distance (FID) score. Furthermore, using this model for data augmentation in image classification tasks effectively mitigates overfitting issues in classifiers.
Key words : generative adversarial networks; data augmentation; image classification

0    引言

視覺深度學(xué)習(xí)的成功不僅僅取決于高容量的模型,,還依賴于大規(guī)模標(biāo)注數(shù)據(jù)的可用性。許多優(yōu)秀的模型在大規(guī)模數(shù)據(jù)集上取得了良好的性能,。然而,,對(duì)于視覺識(shí)別任務(wù),由于數(shù)據(jù)的收集和標(biāo)注耗費(fèi)巨大,,通常在沒有足夠樣本的場(chǎng)景下訓(xùn)練模型,,往往會(huì)導(dǎo)致模型過(guò)擬合,從而降低其泛化性能,。

為了解決這些問題,,數(shù)據(jù)增強(qiáng)是常用的緩解數(shù)據(jù)匱乏的手段之一。雖然傳統(tǒng)的圖像增強(qiáng)技術(shù)(如旋轉(zhuǎn)和隨機(jī)裁剪)的確有效果,,但一些轉(zhuǎn)換可能是無(wú)意義的,,甚至?xí)?dǎo)致圖像語(yǔ)義上的改變。如數(shù)字‘6’旋轉(zhuǎn)180°變成了‘9’,,改變了語(yǔ)義,,這需要專家經(jīng)驗(yàn)進(jìn)行評(píng)估。最近的研究表明,,使用生成對(duì)抗網(wǎng)絡(luò)(GANs)進(jìn)行數(shù)據(jù)增強(qiáng)具有巨大潛力,。生成對(duì)抗網(wǎng)絡(luò)是一種隱式生成模型,,通過(guò)對(duì)真實(shí)樣本的數(shù)據(jù)分布進(jìn)行建模,可以根據(jù)已有數(shù)據(jù)集的高維特征組合生成與訓(xùn)練集完全不同的圖像,,能夠?yàn)樯疃葘W(xué)習(xí)模型提供更多的圖像特征,,以緩解過(guò)擬合問題,。Mirza等人提出的條件生成對(duì)抗網(wǎng)絡(luò)(cGAN)可以通過(guò)控制類別生成對(duì)應(yīng)類別的樣本,,因此可以得到帶有標(biāo)簽的樣本。



本文詳細(xì)內(nèi)容請(qǐng)下載:http://forexkbc.com/resource/share/2000005378




作者信息:

楊鵬坤,,李金龍,,郝潤(rùn)來(lái)

(中國(guó)科學(xué)技術(shù)大學(xué)計(jì)算機(jī)科學(xué)與技術(shù)學(xué)院,安徽合肥230026)


微信圖片_20210517164139.jpg

此內(nèi)容為AET網(wǎng)站原創(chuàng),,未經(jīng)授權(quán)禁止轉(zhuǎn)載,。