Body Mass Index Obesity BMI and Health Assignment

Assignment Task

Abstract

A person's body mass index (BMI) is a gauge of how healthy they are in relation to their weight. Numerous aspects, including physical health, mental health, and popularity, have been linked to BMI. BMI calculations frequently call for precise measurements of height and weight, which entail labor-intensive manual labour. Governments and businesses can employ large-scale automation of BMI calculation to analyse different facets of society and to make smart decisions. Previous research has relied solely on geometric facial features, disregarding all other information, or has employed a data-driven deep learning approach where the volume of data becomes a bottleneck.We employed cutting-edge pre-trained models like Inception-v3, VGG-Faces, VGG19, and Xception and refined them using discriminative learning on the rather sizable public dataset. For training, we used the broader Illinois DOC labelled faces dataset, and for evaluation, we used the arrest records, VIP attribute.

Introduction

The A vital health indicator is a person's BMI (body mass index). If the person is underweight, normal, overweight, or obese, it is determined. Health is currently one of the most overlooked factors. Even technology with many advantages has its downsides. It has made people more slothful, which has decreased their physical activity and resulted in a sedentary lifestyle and an increase in BMI, both of which are harmful to their health and raise the risk of chronic diseases. The likelihood of acquiring cardiovascular and other hazardous diseases increases with increasing BMI. On the other hand, some people struggle with issues like inadequacies and malnutrition. BMI can therefore assist a person in keeping track of their health. 36% of the population, or one in three individuals, are fat on average; by 2030, it is predicted that 20% of people worldwide would be obese.A person can be learned a lot just by looking at their face. Recent research has demonstrated a significant relationship between the BMI of the individual and the human face. People with thin faces are likely to have lower BMIs, and vice versa. Obese people typically have bigger middle and lower facial features. Without a measuring tape and a scale, it is challenging for the person to calculate their BMI. Deep learning has made tremendous strides recently, enabling models to extract useful information from photos. These techniques allow us to extrapolate the BMI from human faces. Therefore, we have suggested a method to predict BMI from human faces in this work. This technique might make it easier for health insurance firms to keep track of their clients' medical histories. Additionally, the government might monitor the health records of a certain area and develop regulations in accordance with them.

Our Contribution

The project “Computing Body Mass Index From a Facial Image using Deep Learning” is a remarkable contribution to the field of healthcare. This innovative project has the potential to revolutionize the way BMI is calculated, as it uses deep learning to accurately measure Body Mass Index from a facial image. This method is more convenient and accurate than traditional methods, as it eliminates the need for manual measurements and provides more accurate results. Additionally, this project can help reduce the burden of healthcare professionals, as they will no longer need to manually measure BMI. This project has the potential to significantly improve the accuracy and efficiency of BMI calculations.The following is the paper's contributions: To establish a link between BMI and facial features of people and to develop a method for predicting BMI from facial features of people utilising deep learning and transfer learning models like VGG-Face, Inception-v3, VGG19, and Xception.

Literature Survey

Laxmi Soni, Ashutosh and Shilpa Datar[2] proposed application of Viola-Jones algorithm with some specific threshold values. The authors come to the conclusion that the Viola-Jones algorithm produces results with a high detection rate and accuracy when a particular threshold value is used. When the image size and resolution are huge and high, the calculation time in this method increases. This work determined that the average detection rate is 97.41%. The presence of many faces in the image has no impact on the calculation time or detection rate.

The computational method was utilised by Wen and Guo [3] to determine BMI. By employing the Active Shape Model to extract facial landmarks from facial photos, the authors were able to extract seven facial traits. Eye size, CJWR (Cheek to Jaw Width Ratio), PAR (Perimeter to Area Ratio), WHR (Width to Upper Facial Height Ratio), FW/FH (lower Face to Face Height Ratio), and MEH are the seven characteristics (Mean Eyebrow height). The regression issue was then solved using the Support Vector Regressor (SVR), Least Square Estimation, and Gaussian Process. . They evaluated and trained using the Morph II dataset. SVR provided the greatest outcomes for both sets of data, according to the results. Barr et al. proposed the estimation of Facial BMI (fBMI) from facial photographs using a similar methodology. For evaluation, there was a stronger association between fBMI and BMI for the normal and overweight categories but a weaker correlation for the underweight and obese groups.Even if the results are significant, only facial landmarks are taken into account when extracting features. By taking more elements out of facial photos, more advancements can be realised.

A customised end-to-end CNN network was suggested by Hera Siddiqui et al. [4] to predict BMI. With the use of pre-trained CNN models including VGG-19, ResNet, DenseNet, MobileNet, and LightCNN, the authors additionally retrieved characteristics from the facial photos and then sent them to SVR and RR for final predictions. With the help of the VisualBmi, VIP attribute, and Bollywood Datasets, they were able to reach Mean Absolute Error (MAE) values between [1.04] and [6.48]. Ridge Regression improved the performance of DenseNet and ResNet models. Pretrained models outperformed the end-to-end CNN model somewhat. Jiang et alanalysis of geometry- and deep learning-based methods for computing visual BMI [5] also covered the impact of several variables like gender, ethnicity, and head orientation on prediction accuracy. Although deep learning-based algorithms outperformed geometry-based ones in terms of performance, the high dimensionality of features has a negative impact because training data is generally scarcer. Large head posture changes also have a negative impact on performance. The authors' FIW-BMI dataset and the Morph II dataset are both derived from social media platforms.

In order to predict a person's BMI from their social media photographs, Kocabey et al. [6] devised a computer vision approach. They used photos from the VisualBMI Project. They utilised 4206 face pictures on VGG-Net and VGGFace models for deep feature extraction. They employed an epsilon for BMI regression.This vector regression model is supported. VGG-Faces outperformed the VGG-Net model. The Pearson correlation coefficient for the test set generated by the VGG-Face model was 0.71, 0.57, and 0.65 for the Male, Female, and Overall categories, respectively. They also demonstrated human vs. machine prediction, with people outperforming machines in lower BMI categories but no difference in higher BMI categories.

Ankur Haritosh et al. [7] suggested an unique approach for estimating height, weight, and BMI from face photographs. They utilized 4206 photographs from the VisualBMI project and 982 images from other sources.

Reddit's HWBMI database. The photos are cropped to 256 256 after running the Voila Jones Face Detection algorithm. These photos are sent into the feature extractor model, which extracts high-level features before feeding them into the 3-layered ANN model. Using the Face to BMI and Reddit HWBMI datasets, XceptionNet MAE for BMI was 4.1 and 3.8, respectively. On the Reddit HWBMI, the MAE was 0.073 for height supplied by VGG-Face and 13.29 for weight given by the XceptionNet model.

Christine Mayer et al. [8] presented a statistical method for determining the association between BMI and waist-to-hip ratio (WHR) and face shape and texture. The authors used the Windows programme TPSDig to mark 119 anatomical landmarks and semilandmarks. They computed semilandmark's exact locations using a sliding landmark technique. Their analysis included 49 standardised pictures of women with BMIs ranging from [17.0, 35.4] and WHRs ranging from [0.66, 0.82]. The form of the face is represented by the Procrustes shape coordinates, and the texture is represented by the RGB values of the standardised photos. The desired associations were tested using the multivariate linear regression approach. BMI was more predictable than WHR from facial traits, with 25% of the variance explained by face shape and 3-10% explained by facial features.

Methodology

Data Preprocessing

The dataset's front-facing images are what we're using. On the other hand, some images have inconsistent zoom levels and tilted head positions. Using the DLIB 68 landmark detection model to align the face vertically and then blurring the background while focusing on the face, we used the preprocessing by StyleGan FFHQ Dataset to make the images similar. Our dataset was converted into Tensorflow Records Dataset (TFRecord) format for training on Tensor Processing Units (TPU). Each record was made up of 1024 images that were preprocessed to be 256 256 3 in size. transfer learning's underlying network. The BMI number appears as the image's label.

Transfer Learning

It can be difficult to calculate BMI using facial images. Therefore, learning all necessary features from comparatively small datasets would be impossible. Transfer learning has been used to improve performance and shorten training times for many computer vision tasks. We have therefore employed cutting-edge pre-trained models like:

• Inception-v3 : A deep convolutional is Inception-v3. network of neurons. It is the Google-developed Inception CNN's third edition model. On what is trained this model. more than a million pictures from the well-known ImageNet. database. It provided a 0 point 779 Top-1 accuracy score and an. Top-5 accuracy rating of 0.937 with roughly 24 Million. only, making it computationally efficient when. in comparison to other models.
• VGG-Face : The developed the VGG-Face model. VGG (Visual Geometry Group) researchers at. Oxford. VGGFace2 [14] is the dataset that this model uses. includes 9131 subjects' facial images, totaling 331 million. It. was developed with the main goal of training robust. face recognition software. A Resnet-50 that had been trained was employed. for our research, a model of this dataset.
• VGG-19 : The deep convolutional VGG-19 model. neural network, the VGG-16 model's replacement, and. developed by Visual Geometry researchers. Group. It is trained on the same data as the Inception-v3 model. well-known ImageNet repository. With Top-1 accuracy, it succeeded. score of 0.752 and a Top-5 accuracy score of 0.925. 143 million parameters, roughly.
• Xception : Xception's feature extraction foundation. 36 convolutional layers comprise the model. It has. extension of the Inception model architecture in which it. uses depthwise Separable Convolutions in place of the. standard inception modules. Additionally, it received training from the ImageNet Database, earning a Top-1 accuracy score. 0.79 and a Top-5 accuracy rating of 0.94 with roughly 22. 1,000,000 parameters.

Training

At the conclusion of everything, we used the same fully connected layers. pre-trained models. Global Average Pooling is used before the output of the pre-trained model is fed into the fully connected layers. To prevent overfitting, we added one dropout layer with a dropout of 50% to our model architecture. The properties of the RELU activation function, Dropout, and Zoneout are all combined in the Gaussian Error Linear Unit (Gelu) activation function that we also used. As a result, we used it in our models because it tends to generalize better when there is more noise in the data. We improved our models as we discovered a relatively larger dataset for our research. The input-proximate layers of deep convolutional networks learn the fundamentals. features like edges and corners. The layers typically learn more complex features from the images used for training as we get closer to the output. In order to increase the amount of features the pre-trained model can extract from the images, we used a higher learning rate for the new fully connected layers and a significantly lower learning rate for some of the final layers. Because of these factors, Adam optimizers were used, and their learning rates were reduced. as we advance through the model's layers. We did so with the aid of TensorFlow Addons' MultiOptimizer.

This Medical Science has been solved by our PHD Experts at My Uni Paper.