2016-5-5 Unnamed files

2016-5-5 Untitled File New template small book maker 1. Abstract of thesis

This paper presents a new discriminativate deep metric learning (ddml) method for face verification in the natural environment. Unlike the existing approach to learning a Markov distance metric to maximize distance between classes, theDdml method trains a deep neural network to learn a series of hierarchical nonlinear Transformations (a set of hierarchical nonlinear Transformations), each face image is projected to the same feature subspace, so that the distance between the positive sample pairs is less than the threshold value, and the negative sample pair distance is greater than the threshold value.

Sfdasd

2. Core approach

DDML flowchart
First, a deep neural network is constructed to transform the image into the network, and the activation function used is Tanh(in the thesis experiment, Tanh is better than sigmoid effect). After mapping the image to a neural network, the output of the top-level network is availableAndIndicates that the distance between them can be calculated using the Euclidean distance:。
We want to explore the differentiated information of the face expression of the top layer of the model, and naturally we hope that the distance between the positive sample pairs is less than the distance between the negative sample pairs.

As shown, before the nonlinear mapping of the network, the distance between the positive sample pairs may be greater than the distance between negative sample pairs, so the learning goal of the neural network is to Search for a nonlinear mapping function, so that the distance between the positive sample pairs is less than $\tau_1$, and the distance between the negative sample pairs.

2016-5-5 Unnamed files