基於特徵金字塔網路之新型異常圖像偵測系統

Abstract

異常資料偵測在神經網路的安全議題上，是一個值得探討的方向。一旦訓練好的神經網路遇上了無法識別的資料型態，就極有可能發生錯誤的判斷，導致無可挽回的後果，像是自動駕駛以及醫學診斷系統就是其中經典的例子。因此，一個有效的分類器不只應該要能準確的識別原先的類別項目，也要能辨識出不屬於他認知範圍的異常資料。本論文提出一個基於特徵金字塔網路之異常圖像偵測系統。相比起其他異常檢測系統的單一輸入單一預測值，我們將系統結合了特徵金字塔網路，因此針對單一影像輸入，可以輸出多尺度的預測值，透過統整多尺度的預測值，有效地讓系統準確度提升。實驗結果顯示此系統不僅可以保留原先任務需求，且在多個視覺資料集上皆顯示辨識效果有所提升。Anomaly data detection is a direction worthy of discussion on the security issue of neural networks. Once an unrecognized data type is input into a trained neural network which is very likely to make wrong judgments, and it will lead to an irreversible consequences, such as autonomous driving and medical diagnosis systems are classic examples.In this paper, we propose an out-of-distribution image detection system based on feature pyramid network. Compared with the single input and single prediction of other out-of-distribution detection systems, we combine the system with the feature pyramid network, so for a single image input, multi-scale prediction would be output. Experimental results show that the system not only retains the original task, but also improves the recognition accuracy on multiple visual datasets.