KomNET: Face Image Dataset from Various Media for Face Recognition

KomNet is a face image dataset originated from three media sources which can be used to recognize faces. KomNET contains face images which were collected from three different media sources, i.e. mobile phone camera, digital camera, and media social. The collected face dataset was frontal face image or facing the camera. The face dataset originated from the three media were collected without certain conditions such as lighting, background, haircut, mustache and beard, head cover, glasses, and differences of expression. KomNet dataset were collected from 50 clusters in which each of them consisted of 24 face images. To increase the number of training data, the face images were propagated with augmentation image technique, in which ten augmentations were used such as Rotate, Flip, Gaussian Blur, Gamma Contrast, Sigmoid Contrast, Sharpen, Emboss, Histogram Equalization, Hue and Saturation, Average Blur so the face images became 240 face images per cluster. The author trained the dataset by using CNN-based transfer learning VGGface. KomNET dataset are freely available on https://data.mendeley.com/datasets/hsv83m5zbb/2.

© 2020 The Author(s).Published by Elsevier Inc.This is an open access article under the CC BY license.

Value of the Data
• KomNET has face images originated from mobile phone camera, digital camera, and media social; • KomNET can be used for training, validation, and algorithm comparison for face recognition.
• Dataset KomNET originated from the three media were collected without certain conditions such as lighting, background, haircut, mustache and beard, head cover, glasses, and differences of expression.The number of data training was increased by using augmentation image technique; • KomNET can be used to develop the new CNN-based transfer learning architecture or modifying the existing architecture, e.g.Stochastic Gradient Descent or ImageNet, to improve layer efficiency on face recognition.• Researchers who are researching about facial recognition can use this KomNET.

Data Description
KomNET dataset images contains more than 39,600 face images originated from mobile phone camera, digital camera, and media social.The purpose is training, validating and recognizing face with CNN-method, or other technique.The use of dataset for face recognition usually uses images of photos originated from single media such as dataset from mobile phone [1 , 2] , Facebook [3] , digital camera [4 , 5] .Algorithm development for face recognition requires images dataset from various media sources, it is a challenge for researchers because the expected results

Experimental Design, Materials, and Methods
KomNET images were collected from three various sources, i.e. mobile phone camera, digital camera, and media social.The collection of face images was collected with frontal face which facing camera without considering background, lighting, expression, glasses, head cover, etc.Every collected image from the three sources and image training process was separated into folder train and folder test.The collected images from these three sources have different sizes, therefore there was re-sizing process to make the size the same of 224 × 224 pixel.For good result in face recognition, there should be lots of training, and if the data is few so more data is needed and there will be minor changes in dataset.The change can be done by changing the face image such as translation, rotate, or viewpoint, size or illumination or the combination and this way can be done with image augmentation technique.The used augmentation images on dataset KomNET were average blur, emboss, flip, gamma contrast, gaussian blur, histogram equalization, rotate, hue and saturation, sharpen, and sigmoid contrast.After augmentation, the images were inserted into folder augmentation.The example of image after augmentation is presented as the following Fig. 1 .
This dataset has also been used in face recognition using the CNN algorithm.Image from social media has received approval from the owner of the relevant social media account.Dataset testing was done in Computer Laboratory, Department of Electrical Engineering, Politeknik Negeri Bali, Bali, Indonesia.For the initial process, the author used wavelet method to get face feature.Furthermore, the face feature was processed by CNN-based transfer learning VGG face.The researcher [9] said that first layer feature is general and the last layer feature is specific so for the last 2 layer there was no training, only the result of feature was taken as the output.Next, this feature was processed or done by fine tuning with classification model.
Transfer learning is a good method in computer vision because it is so accurate and saving time in building a model [10] .It is because transfer learning can be used to solve similar problem without starting learning process from the beginning, by improving previous learning.Researcher [11] said that picture classification problem on deep learning can be solved through good transfer learning.
Transfer learning in computer vision is usually expressed through pre-trained model.Model pre-trained is a model usually used for training big dataset in solving similar problem.Therefore, by computation consideration for these model training, usually researchers import and use model from published literature (e.g.VGG, Inception, MobileNet).The author used pre-trained model which was based on convolutional neural networks (CNN) as conducted by researcher [12] .