Gender Analysis Using Simplified Local Binary Pattern Features Under Uncontrolled Environment

Recently in Malaysia, digital out-of-home (DOOH) advertising is getting common. It is a marketing tool to promote the branding of a product or service in the high population flow area. The existing DOOH advertising in Malaysia only broadcast the advertisement in a continuous and random manner withou...

全面介绍

Saved in:
书目详细资料
主要作者: Loo, Eng Keong
格式: Thesis
出版: 2021
主题:
标签: 添加标签
没有标签, 成为第一个标记此记录!
实物特征
总结:Recently in Malaysia, digital out-of-home (DOOH) advertising is getting common. It is a marketing tool to promote the branding of a product or service in the high population flow area. The existing DOOH advertising in Malaysia only broadcast the advertisement in a continuous and random manner without considering the presence of the audience. This causes the audience most likely to perceive some irrelevant advertisements. The recent advances of facial analysis unlock the potential of the DOOH media towards the targeted advertising manner. Therefore, the DOOH media is able to deliver a customised advertisement based on the facial attributes of the audience. However, the DOOH media is commonly operating in an uncontrolled environment. In such an environment, facial analysis faces two main challenges which are blurry faces and misaligned faces. These challenges are noises that increase the difficulty in feature extraction. In this thesis, meFusion, a handcrafted feature extraction method for gender estimation is proposed to handle the aforementioned challenges without going through the pre-processing stage. Bypassing the preprocessing stage is able to reduce the workload of DOOH media. A series of experiments that simulate the challenges are conducted. The experiments result show meFusion has the highest tolerance towards the noises, while compared with the existing handcrafted feature extraction method. A designed gender estimation framework that incorporates meFusion to realise the targeted advertising on DOOH media. A series of experiments that compare the designed framework with a pre-trained model from cloud services such as Google Cloud’s Vision API and Microsoft Azure Face. At the same time, a newly collected Malaysian face dataset in an uncontrolled environment, named Malaysian Ethnics Facial Database 2 (MEFD2), which consist of 2975 persons with a total of 13498 images, is applied in the experiments.