In this paper a novel multikernel deterministic extreme learning machine (ELM) and its variants are developed for classification of non-linear problems. Over a decade ELM is proved to be efficacious learning algorithms, but due to the non-deterministic and single kernel dependent feature mapping proprietary, it cannot be efficiently applied to real time classification problems that require invariant output solution. We address this problem by analytically calculation of input and hidden layer parameters for achieving the deterministic solution and exploiting the data fusion proficiency of multiple kernel learning. This investigation originates a novel deterministic ELM with single layer architecture in which kernel function is aggregation of linear combination of disparate base kernels. The weight of kernels depends upon perspicacity of problem and is empirically calculated. To further enhance the performance we utilize the capabilities of fuzzy set to find the pixel-wise coalition of face images with different classes. This handles the uncertainty involved in face recognition under varying environment condition. The pixel-wise membership value extracts the unseen information from images up to significant extent. The validity of the proposed approach is tested extensively on diverse set of face databases: databases with and without illumination variations and discrete types of kernels. The proposed algorithms achieve 100% recognition rate for Yale database, when seven and eight images per identity are considered for training. Also, the superior recognition rate is achieved for AT & T, Georgia Tech and AR databases, when compared with contemporary methods that prove the efficacy of proposed approaches in uncontrolled conditions significantly.

在本文中，开发了一种新颖的多核确定性极限学习机 (ELM) 及其变体，用于非线性问题的分类。十多年来，ELM 被证明是有效的学习算法，但由于非确定性和单内核相关特征映射专有，它无法有效地应用于需要不变输出解决方案的实时分类问题。我们通过分析计算输入和隐藏层参数来实现确定性解决方案并利用多核学习的数据融合能力来解决这个问题。这项研究起源于一种具有单层架构的新型确定性 ELM，其中核函数是不同基核的线性组合的聚合。内核的权重取决于问题的清晰度，并根据经验计算。为了进一步提高性能，我们利用模糊集的功能来找到不同类别的人脸图像的像素级联盟。这处理了在不同环境条件下人脸识别所涉及的不确定性。像素级隶属度值在很大程度上从图像中提取了看不见的信息。所提出的方法的有效性在不同的人脸数据库集上进行了广泛的测试：有和没有光照变化和离散类型的内核的数据库。所提出的算法对耶鲁数据库实现了 100% 的识别率，当每个身份的 7 和 8 张图像被考虑用于训练时。此外，对于 AT&T、Georgia Tech 和 AR 数据库实现了卓越的识别率，