跨域掌紋辨識的資料擴增

Abstract

基於生物特徵信息的身份認證系統在近年被廣泛的運用，大多數人已經接受它帶來的便利性、安全性和隱私性，可以避免使用者的密碼被盜或忘記。其中掌紋識別是人性化且衛生的生物識別，無需觸摸設備，對手勢也沒有太大的限制。隨著當今影像識別技術的飛速發展，深度學習方法在多樣化的影像資料處理上表現優於傳統方法，然而大多數掌紋識別中的深度學習方法僅在單個資料集上進行訓練和測試，一旦使用不同的成像設備，辨識的效果就會嚴重受到影像。但在實際應用場景中，輸入影像來自不同的相機是很常見的，因此如何解決不同成像設備帶來的影像差距至關重要。因此我們提出了三種影像增強策略和 ResNeSt 的縮減版本來解決跨域的掌紋識別。首先我們在訓練期間使用 Optuna 框架和 TPE 採樣器對影像隨機轉換進行超參數優化搜索。其次，將 ROI 影像旋轉到四個方向作為不同類別的訓練影像來 oversample 訓練資料集。這也可以用於增強我們提出的基於 test-time augmentation 的多轉換特徵比對方法。在有約束的資料集 (PolyU-M) 上訓練並在無約束的資料集 (MPD) 上測試的困難條件下，三種增強方法可以分別提高準確度 12.55%、5.34% 和 4.66%，總共可以實現 22.55% 的準確度提升。 此外，使用 MPD 訓練的模型在所有測試資料集中都可以達到 99.66% 以上的準確度。

Identity authentication systems based on biometric information have been widely used in recent years. Most people accept its convenience, security, and privacy, avoiding the user's password being stolen or forgotten. Among them, palmprint recognition is user-friendly and hygienic biometric, without touching the device and too many pose restrictions. With the rapid development of today's image recognition technology, deep learning methods perform better than traditional methods on diverse image data. Although most deep learning methods in palmprint recognition perform well on a single dataset, it will seriously affect performance once applied to different image acquisition devices. In practical application scenarios, it is common for input images to come from different cameras, so how to bridge the gap between different imaging conditions is extremely crucial. Therefore, we propose three data augmentation strategies and a reduced version of ResNeSt to solve the cross-domain palmprint recognition problem. First, we perform a hyperparameter optimization search for random transformations during training with the Optuna framework and the TPE sampler. Second, the training dataset is oversampling augmented, which the ROI images are rotated into four orientations as training images of different classes. This can also be used to enhance our proposed multi-transform matching based on a test-time augmentation technique. Under the difficult conditions of training on a constrained acquisition dataset (PolyU-M) and testing on an unconstrained dataset (MPD), the three augmentation methods can improve by 12.55\%, 5.34\%, and 4.66\%, respectively, so a total of 22.55\% improvement can be achieved. Furthermore, the model trained on both MPD can achieve more than 99.66\% accuracy in all test datasets.