基於複合損失函數於強健注意力卷積自動編碼器於低品質指紋圖像重建;Multi-loss Function in Robust Attention Convolutional Autoencoder for Reconstruction low-quality Fingerprint Image

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/89799

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/89799

题名:	基於複合損失函數於強健注意力卷積自動編碼器於低品質指紋圖像重建;Multi-loss Function in Robust Attention Convolutional Autoencoder for Reconstruction low-quality Fingerprint Image
作者:	哈弗嵐;Halberd, Franki
贡献者:	資訊工程學系
关键词:	指紋圖像;重建;多重損失函數;Fingerprint Image;Reconstruction;Multi Loss Function
日期:	2022-07-25
上传时间:	2022-10-04 12:00:18 (UTC+8)
出版者:	國立中央大學
摘要:	指紋已被廣泛用於身份驗證和驗證。除嚴重損傷外，解剖方法更穩定且不可改變[1]。季節性喜歡低溫和乾燥的地區通常會導致皮膚變得乾燥。因此，即使用指尖正確觸摸指紋傳感器的表面也會產生幹指紋。然而，COVID-19 大流行（一種於 2019 年爆發的冠狀病毒疾病）已迅速發展成為一個重要的全球公共衛生問題，這促使人們更廣泛地使用洗手液。被迫定期用消毒劑洗手的人也會留下乾燥的指紋 [2]。即使正確觸摸指紋傳感器的表面，這種情況也可能會影響獲取的指紋圖像的質量。當指紋不完美時，從指紋中提取細節比完整時更困難，並且得到錯誤細節的可能性更高。在我們的研究中，我們提出了基於具有軟注意力的捲積神經網絡自動編碼器重建指紋圖像的不完整性。我們將圖像質量評估（IQA）方面的感知測量作為損失函數來提供足夠的權重校正。我們設計了幾種損失函數的組合來研究最佳模型，例如多尺度結構相似性指數度量 (MS-SSIM)、結構相似性指數度量 (SSIM) 和峰值信噪比 (PSNR)，均值平方誤差 (MSE) 和平均絕對誤差 (MAE)。我們從實驗結果中獲得最高的圖像質量度量分數，總結為我們提出的方法作為損失函數（SSIM + MSE）和優化器均方根傳播（RMSProp）。我們使用來自深度學習和媒體系統實驗室的 3 人的 85 張指紋圖像評估了圖像重建。最終，我們提出的方法獲得了令人印象深刻的結果，將圖像的平均質量提高了 34.02%，SSIM 提高了 10.13%，MSE 提高了 64.38%，FSIM 提高了 10.56%。在定量指標和人類視覺判斷方面表現良好。我們提出的方法可以作為未來研究的基準，以提高識別準確度的性能，特別是在低質量指紋問題上。 ;Fingerprint has been prominently used for authentication and verification of a person. The anatomical approaches are more stable and non‐alterable, except by severe injury [1]. Seasonality likes low temperature and dry region commonly leads the skin become dry. As a result, even correctly touching the surface of a fingerprint sensor with your fingertip can generate a dry fingerprint. However, the COVID-19 pandemic (a coronavirus illness that broke out in 2019) has quickly grown to be a significant worldwide public health issue, which has prompted a wider use of hand sanitizers. People who are compelled to routinely wash their hands with disinfectant also get dry fingerprints [2]. This condition may affect the quality of the acquired fingerprint image even correctly touching the surface of a fingerprint sensor. When a fingerprint is imperfect, it is more difficult than when it is complete to extract the minutiae from it, and the likelihood of getting the wrong minutiae is higher. In our research, we proposed reconstructing the incompleteness of fingerprint images based on convolutional neural network autoencoders with soft attention. We incorporate the perceptual measurement in terms of Image Quality Assessment (IQA) as the loss function to give adequate weight correction. We design the combination of several loss functions to investigate the best model, such as the multiscale structural similarity index measure (MS-SSIM), the structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR), mean square error (MSE), and mean absolute errors (MAE). We obtain the highest image quality metric scores from the experimental result summarized as our proposed method, which is a loss function (SSIM + MSE) with optimizer Root Mean Squared Propagation (RMSProp). We evaluated the image reconstruction using 85 fingerprint images from 3 persons from Deep Learning and Media System Lab. Eventually, our proposed method gets impressive results, increasing the image′s average quality by PSNR of 34.02%, SSIM of 10.13%, MSE of 64.38%, and FSIM of 10.56%. The performance is good in terms of quantitative metrics and human visual judgment. Our proposed method can be the baseline for future research to increase the performance in recognition accuracy, especially in the problem of low-quality fingerprints.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	73	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....