Abstract:

Deep Neural Network （DNN） may encounter samples that are not present during training， and there are difficulties in rejecting unknown class samples accurately. Open set recognition can classify known class samples accurately while rejecting unknown class samples. In the current domain of open set recognition， prototype learning methods have gained widespread applications. However， these methods cannot ensure intra-class compactness and inter-class separability of sample distribution simultaneously. Therefore， an open set recognition method with Open Generation and Feature Optimization （OGFO） was proposed. Firstly， the concept of open point was presented， and the inherent features of corresponding class samples were learned by prototype points through DNN. The average of prototype points for each class was open point， representing the inherent features of unknown classes， and the central region of the feature space was occupied by the open point as open space for the distribution of unknown class samples. Then， a Feature Optimization Algorithm based on open points （FOA） was proposed， so that open points were utilized to force a more compact distribution within the same class samples， and a more separable distribution among different class samples. Finally， an open point based generation method OGAN （Open Generative Adversarial Network） was proposed， and the unknown class samples generated by OGAN were forced by DNN to distribute in the open space occupied by open points. Experimental results demonstrate that compared with Adversarial Reciprocal Points Learning for open set recognition （ARPL） on datasets such as MNIST， SVHN， CIFAR10， and TinyImageNet， OGFO has significant improvements in AUROC （Area Under the Receiver Operating Characteristic curve）. Especially， on TinyImageNet dataset， OGFO improves the AUROC by at least 3 percentage points and has improvements of at least 6 and at least 5 percentage points， respectively， in accuracy and OSCR （Open Set Classification Rate） compared to ARPL. It can be seen that the challenge which other methods cannot address： balancing intra-distribution compactness and inter-distribution separability of sample distribution simultaneously is addressed by OGFO.

Key words: feature optimization, open set recognition, open point, prototype learning, Deep Neural Network (DNN), generator

摘要：

当深度神经网络（DNN）遇到训练时未遇见的类别的样本时，不能准确地拒绝未知类样本，而开集识别能在准确分类已知类样本同时拒绝未知类样本。目前在开集识别领域，原型学习方法广为应用，然而这些方法都无法同时保证样本分布内的紧凑性和样本分布间的分离性。因此，提出开放生成与特征优化的开集识别方法（OGFO）。首先，提出开放点的概念，原型点通过DNN学习对应类别样本的固有特征而开放点是各类别原型点的均值。开放点代表未知类的固有特征且占据特征空间的中心区域。特征空间中心区域为未知类样本分布的开放空间；其次，提出基于开放点的特征优化算法（FOA），从而利用开放点强迫相同类别样本内部的分布更加紧凑并且迫使不同类别样本间的分布更加分离；最后，提出基于开放点的生成方法OGAN（Open Generative Adversarial Network），并使用DNN迫使OGAN生成的未知类样本分布在开放点占据的开放空间中。实验结果表明，相较于基于对抗性反向点学习的开集识别方法（ARPL），OGFO在MNIST、SVHN、CIFAR10和TinyImageNet数据集上的AUROC（Area Under the Receiver Operating Characteristic curve）提升明显，尤其在TinyImageNet数据集上的AUROC上至少提升了3个百分点，在准确率和OSCR（Open Set Classification Rate）上分别至少提升6和5个百分点。可见，OGFO解决了其他方法无法兼顾样本分布内的紧凑性和样本分布间的分离性的问题。

关键词: 特征优化, 开集识别, 开放点, 原型学习, 深度神经网络, 生成器