Gender Classification based on Convolutional Neural Network with Spatial Filtering

• Main Authors: Chen-Shin Li, 李政勳
• Other Authors: Wen-Shiung Chen
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/qq6xe8

碩士 === 國立暨南國際大學 === 電機工程學系 === 105 === Artificial Intelligence originated in 1943, in recent years, gradually walks on the trail, from a neuron mathematical model evolved into multi-layer and complex neural networks, the biggest pusher is Geoff Hinton. In 2006s, Geoff Hinton successfully trained the multi-layer neural networks, and named it deep learning. In 2012, Geoff Hinton led the team to participate in the ImageNet image recognition competition, won the championship with a 16.42% error rate, and brought the trended of deep learning. Traditional face recognition extracts features by collecting data and analyzing data manually, but deep learning is through learning to extract features. In the age of big data, it is possible to find that the feature extraction by manually is not appropriate, because it is time consuming and it is laborious, when used to train the multi-layer model. When deep learning trains the multi-layer model, can learn enough to represent the characteristics of the task, save time and automatically learn the characteristics, that are the deep learning's charm. There are many different structures of models in deep learning. This thesis uses convolutional neural networks of deep learning, because the structure of the convolutional neural networks has a two-dimensional form, that is very suitable for processing images. The convolution neural networks have invariance and weight sharing, so that the parameters quantity of the neural network will not be massive and get the characteristics of invariance. It is a challenge to find the appropriate parameters by adjusting the parameters of convolutional neural networks. Therefore, this thesis proposes a method of combining the convolutional neural networks with spatial filtering, and it is expected that the feature maps after convolution can be robust by passing through the spatial filter, and the model's training will be smoother.