基於半監督式生成對抗網路架構作遙測影像異常檢測及分類

異常檢測在各個領域裡扮演著重要的角色，人們通常對感興趣的資料使用異常檢測方法，得以發現不在預期內的事件或行為。由於台灣特殊的地理環境、夏季颱風和強西南氣流帶來的強降雨量，陡峭的山坡地上常常發生土石流和山崩的事件。為了公眾安全，我們提出一個兼具異常分數計算以及分類能力的異常檢測方法，以此提早防患未然避免重大的損失。在本研究中，我們提出一個結合半監督式學習的對抗生成網路( Generative Adversarial Networks, GAN )和分類功能的異常檢測方法，稱為 GSAC( GAN-based Semi-supervised Anomaly detection and Classification )。透過對抗學習之半監督式學習的生成模型，具有能夠生成樣本和不太依靠標記資料的特色，其架構擴展性佳已廣泛應用於各個領域。而我們提出的方法在面對上述問題時，能夠讓模型在維持著懸殊差距的樣本量體下，合理的學習到正常樣本以及異常樣本的特徵。本研究設計的網路讓生成對抗網路的辨別器同時也擁有分類器的效果，藉此在找出異常樣本的同時，能夠區分異常的類別。我們使用數個主流的資料集包含 MNIST、CIFAR10、UC Merced Land-Use、NWPURESISC45…來驗證我們所提出方法的可靠性。在 MNIST 和 CIFAR10 上，能夠維持平均高於其他傳統方法的分數下，保持其潛在空間並且獲得異常分類的能力。在遙測影像 UC 資料集的分類下，分類結果達到 89%。此結果證明我們的模型不僅能夠依靠原有的異常分數進行異常檢測，也能利用分類結果更精準的發掘出異常事件。我們探討了異常樣本的更新機制，整體來看，在分開更新機制下的 AUC、分類準確率以及正常樣本生成圖片的品質皆比全部更新機制下的數據還要好，而分開更新機制下異常樣本的圖片還原，也比全部更新機制下的圖片還原更能表達出若是異常樣本則無法成功的被還原。研究顯示此模型能夠有效的學習正異資料的特徵及分布，並從異常檢測的分數及異常分類結果的兩種角度，利用不同層面找出異常點。

關鍵字

異常檢測；異常分類；半監督式學習；遙測影像；對抗生成網路

並列摘要

Anomaly detection plays an important role in every field. It’s usually used to discover whether there are unexpected events or behaviors in the observed data. Due to Taiwan’s special topography, typhoons and strong southwesterly flow bring heavy rainfall in summer. there are often landslides on many steep mountain slopes. For public safety, we proposed a method which can identify reclamation regions as soon as possible. Moreover, it will be helpful if we can identify the type of anomalies. In this thesis, we proposed an anomaly classification model called GSAC( GAN-based Semi-supervised Anomaly Classification ) which utilize the generative model and semisupervised learning. . We also propose a learning strategy that could learn to separate anomalous samples while maintaining the discriminability of normal samples in the latent space. With such design, the difference between normal and abnormal samples can be effectively classified. We verify the reliability of our proposed method on different dataset, MNIST, CIFAR10, UC, RESISC45, in order to prove the ability of our method. On the MNIST and CIFAR10 datasets, the model is able to maintain an average score higher than other traditional methods on CIFAR10 and MNIST, and also achieve the 89% accuracy of classification on UC dataset. The evidence proves that our method could not only rely on the score of anomaly detection, but exploit the outcome of the classification for discover abnormal events more accurately. The mechanism for updating anomalous samples was discussed. Overall, the separate update mechanism is preferable to the complete update mechanism on the classification accuracy, AUC, and image quality generated by normal samples. The image restoration of abnormal samples under the separate update mechanism, which can express that an abnormal sample cannot be restored successfully, but the complete update mechanism is unable to express it. We also explore the robustness of the model on CIFAR10 and aerial imagery dataset, research shows that our model could learn the characteristics and distribution of data effectively, and catch out the abnormal dot in difference aspect.

並列關鍵字

Anomaly Detection ； Anomaly Classification ； Semi-supervised Learning ； Remote Sensing Images ； Generative Adversarial Network

參考文獻

[1] N. Sarafijanovic-Djukic and J. Davis, “Fast distancebased anomaly detection in images

Google Scholar

using an inceptionlike autoencoder,” in International Conference on Discovery Science,

Google Scholar

Springer, pp. 493–508, 2019, arXiv:2003.08731.

Google Scholar

[2] T. Lillesand, R.W. Kiefer, J. Chipman, “Remote sensing and image interpretation,” 7th ed.,

Google Scholar

John Wiley & Sons, 2015.

Google Scholar

主題瀏覽