مقاله انگلیسی رایگان در مورد انطباق دامنه عمیق بررسی نشده برای تشخیص چهره – IEEE 2018

With availability of a massive amount of labeled images gathered from Internet, deep neural networks have significantly improved the performance in many computer vision applications, such as object detection [3], [15], object classification [7], [17], face recognition [11], [19], [16], [8], [9] and so on. Most of typical techniques are to train a deep neural network with massive images and then apply it to target test dataset. These methods are valid when the training data and test data are independently and identically drawn from the same or similarity distribution. However, in actual application scenario, the distribution of target and training data is always dissonant, which degenerates model performance on target test data. To accommodate the distribution of target data and enhance model performance, one of the most direct approach is to fine-tune a pre-trained deep neural network’s parameter on target database with the supervision of data label. This strategy turns out to be problematic for a target task where labeled data is lacking or even unavailable. Meanwhile, the deep neural networks easily suffer from a significant amount of over-fitting under supervision of small amount of labeled data, which usually degenerates the generalization ability of the model. This problem also exists in the face recognition task. Most of deep models such as VGG-face [11], Facenet [16], SphereFace [8] et al. , are firstly trained on a large scale face database and then evaluated on other face databases like LFW [4], Youtube [20], MegaFace [6] et.al. These databases are all gathered from Internet for the convenience of data collection. However, in real world applications, the target test data may contain people with specific ethnicity, age group, gender, imaging quality, pose of faces etc. and the shooting environment of target test data and source training data may vary greatly. These factors increase domain discrepancy and degenerate face recognition performance on target application. As shown in Fig. 1, there are significant differences in the picture of different databases. The images in CASIA-WebFace [22] are collected from Internet under unconstrained environment and most of the figures are celebrities and public. The GBU [13] contains still frontal facial images acquired with digital camera. FERET [14] is collected under constrained environment and the pictures are all gray. Different data collection methods and application environments cause a significant discrepancy between different databases. Our experiment results also show a poor performance on target test database when directly adopting the pre-trained model.