Average biased ReLU based CNN descriptor for improved face retrieval

The convolutional neural networks (CNN), including AlexNet, GoogleNet, VGGNet, etc. extract features for many computer vision problems which are very discriminative. trained CNN model over one dataset performs reasonably well whereas on another of similar type the hand-designed feature descriptor outperforms same model. Rectified Linear Unit (ReLU) layer discards some values in order to introduce non-linearity. In this paper, it is proposed that discriminative ability deep image representation using can be improved by Average Biased ReLU (AB-ReLU) at last few layers. Basically, AB-ReLU improves two ways: 1) exploits and discarded negative information 2) also neglects irrelevant positive used ReLU. VGGFace MatConvNet VGG-Face as face retrieval other datasets. approach tested six challenging, unconstrained robust datasets (PubFig, LFW, PaSC, AR, FERET ExtYale) a large scale (PolyUNIR) framework. It observed when with pre-trained validation error training network after replacing all ReLUs AB-ReLUs favorable each dataset. even state-of-the-art activation functions, such Sigmoid, ReLU, Leaky Flexible seven