Exploring Transfer Learning on Face Recognition of Dark Skinned, Low Quality and Low Resource Face Data
EExploring Transfer Learning on Face Recognition ofDark Skinned, Low Quality and Low Resource FaceData
Nuredin Ali
Department of Information SystemsMekelle University [email protected]
Abstract
There is a big difference in the tone of color of skin between dark and light skinnedpeople. Despite this fact, most face recognition tasks almost all classical state-of-the-art models are trained on datasets containing an overwhelming majority of lightskinned face images. It is tedious to collect a huge amount of data for dark skinnedfaces and train a model from scratch. In this paper, we apply transfer learning onVGGFace to check how it works on recognising dark skinned mainly Ethiopianfaces. The dataset is of low quality and low resource. Our experimental resultsshow above 95% accuracy which indicates that transfer learning in such settingsworks.
Face recognition (FR) is a technology capable of identifying or verifying a person from a digitalimage or a video frame. [1] Face recognition has been a prominent bio-metric technique for identityauthentication and has been widely used in many areas such as military, finance, public security,and everyday life. Most of the classical state-of-the-art models are trained on very large datasetsof mostly light skinned faces. Most of the people in African countries have dark skinned faces andcurrently there are no readily available datasets collected for researchers to make such experiments.It is tedious to collect a huge amount of data and train a model from scratch. The most efficienttechnique to use in the case of a low resource is to transfer the knowledge a model has learned onanother data. [2] Transfer Learning is a Machine Learning technique whereby a model is trained anddeveloped for one task and is then re-used on a second related task. In this work, we evaluate howtransfer learning from a model pre-trained on mostly light skinned faces works to recognize a verylow quality and low resource dataset of dark skinned faces.
Research in computer vision has included work on issues that have direct social impact, such assecurity and privacy. However, research on the related issue of diversity and inclusion in vision issurprisingly lacking [3]. The work by [3] focused on gender classification and face detection. Whilein this paper we focus on recognition of individuals by applying transfer learning. The ChaLearn“Looking at People” challenge from [4] provides the Faces of the World (FotW) dataset, whichannotates gender and the presence of smiling on faces. [5] won first place in this challenge, utilizingmulti-task learning (MTL) and fine-tuning on top of a model trained for face recognition [6]. [7] laterpublished an out-performing result for the same task on FotW utilizing MTL and transfer learningfrom a face recognition model. In this case, we use transfer learning to recognize dark skinned facesfrom a model pre-trained on mostly light skinned faces.
Black in AI workshop at the 34th Conference on Neural Information Processing Systems (NeurIPS 2020). a r X i v : . [ c s . C V ] J a n Data and methodology
To develop the dataset for this experiment, 15 students coming from a diversified part of Ethiopiaparticipated. A total of 1,500 images were used (100 for each individual). Figure 1 shows exampleimages from our dataset. 70% of the data is used for training the model and the remaining 30% isused to validate the trained model. The images are collected using a very low-quality camera whichis 0.98MP (megapixels). The data has been collected in a controlled environment. Which can beapplicable to Electronic Gate for instance.First we trained a model from scratch by only having the structure of some of the classical modelslike LeNet and AlexNet. After looking at the results they were not satisfactory. The results arestated below. We used a model pre-trained on a huge dataset of mostly light skinned faces which isVGGFace. The model was trained on VGGFace dataset, a very large-scale dataset 2.6M images, over2.6K people [6]. Figure 1 shows example images from this dataset. While applying transfer learning,Feeding the extracted features as input to a fully connected layer and softmax activation providesbetter result [8].Our experimental settings are as follows. The extracted features are fed in to a fully connected layer.As our experiment, Finetuning deeper results reduction in accuracy as there is limited data to train on.To learn some extra features, Maxpooling, average pooling, dense layer and dropout layers are added.A very low learning rate of 0.001, batch size of 32, activation of softmax, loss function of categoricalcross-entropy and Adam as an optimizer were used to train the face recognition model.Figure 1: Sample of the VGGFace datasetFigure 2: Sample of the dataset used to develop our model
The evaluation metric used in this experiment is accuracy. For each image, we check if the correctlabel is found. VGGFace achieved 98.95% accuracy when it was first developed [6]. Using ourdataset the architecture of LeNet achieved 68% and AlexNet 82%. The model developed using thetransfer learning achieved more than 95% accuracy. This indicates that it is possible to develop amodel by transfer learning from the state-of-the-art VGGFace model.
In this work, we showed experimentally and got an indication that using transfer learning on VGGFaceto recognize a low quality and low resource dark-skinned face data works. This is very promising asit is very tedious to collect a huge amount of data for dark skinned faces and develop a model that hasa high accuracy from scratch. For future works, We encourage vision researchers to explore moretowards such techniques and add on how to make such methods more efficient.2 eferences [1] Wang Mei and Weihong Deng. Deep face recognition: A survey. arXiv preprint arXiv:1804.06655 , 2018.[2] Mahbub Hussain, Jordan Bird, and Diego Faria. A study on cnn transfer learning for imageclassification. 06 2018.[3] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities incommercial gender classification. In
Conference on fairness, accountability and transparency ,pages 77–91, 2018.[4] Sergio Escalera, Mercedes Torres Torres, Brais Martinez, Xavier Baró, Hugo Jair Escalante,Isabelle Guyon, Georgios Tzimiropoulos, Ciprian Corneou, Marc Oliu, Mohammad Ali Bagheri,et al. Chalearn looking at people and faces of the world: Face analysis workshop and challenge2016. In
Proceedings of the IEEE Conference on Computer Vision and Pattern RecognitionWorkshops , pages 1–8, 2016.[5] Kaipeng Zhang, Lianzhi Tan, Zhifeng Li, and Yu Qiao. Gender and smile classification usingdeep convolutional neural networks. In
Proceedings of the IEEE Conference on Computer Visionand Pattern Recognition Workshops , pages 34–38, 2016.[6] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. 2015.[7] Rajeev Ranjan, Swami Sankaranarayanan, Carlos D Castillo, and Rama Chellappa. An all-in-oneconvolutional neural network for face analysis. In , pages 17–24. IEEE, 2017.[8] R. M. Prakash, N. Thenmoezhi, and M. Gayathri. Face recognition with convolutional neuralnetwork and transfer learning. In2019 International Conference on Smart Systems and InventiveTechnology (ICSSIT)