FACE IMAGE SUPER RESOLUTION USING MACHINE LEARNING

High-resolution images are usually desired in digital imaging applications for later image processing. Image resolution describes the details of an image, the higher the resolution and the more details of an image. Super-resolution (SR) are techniques that construct high-resolution (HR) images from

Project Title

FACE IMAGE SUPER RESOLUTION USING MACHINE LEARNING

Project Area of Specialization

Artificial Intelligence

Project Summary

High-resolution images are usually desired in digital imaging applications for later image processing. Image resolution describes the details of an image, the higher the resolution and the more details of an image. Super-resolution (SR) are techniques that construct high-resolution (HR) images from several observed images, thereby increasing the high frequency components and removing the degradations caused by the imaging process of the low resolution camera. The basic idea behind Super Resolution is to combine the non-redundant information contained in low-resolution frames to generate a high-resolution image.

Single Super-resolution image (SR) is a classic computer vision issue. This dilemma is fundamentally unfounded since for any given low-resolution pixel there is a multiplicity of solutions. In other words, it's an undetermined inverse problem, the solution of which isn't unique. Recent state-of - the-art methods for super-resolution of single images are mostly based on examples. Such methods either manipulate internal similarities of the same image or learn from external low-and high-resolution exemplar pairs to map functions.

Consider a convolutionary neural network that learns to map end-to-end between low-and high-resolution images directly. This method radically differs from current conventional example-based methods, in that this method does not directly study the dictionaries or multiples for patch space modeling. Implicitly these are done by means of hidden layers. In addition, the selection and aggregation of patches are also formulated as convolutionary layers, so they are involved in the optimization. By learning, the whole SR pipeline is completely obtained in this system, with little pre /post processing. Let’s call the proposed Convolutionary Neural Network (SRCNN) Super-Resolution model. The proposed SRCNN has several appealing properties.

Project Objectives

The key objective of super-resolution (SR) imaging is to reconstruct a higher-resolution image based on a set of images, acquired from the same scene and denoted as ‘low-resolution’ images, to overcome the limitation and/or ill-posed conditions of the image acquisition process for facilitating better content visualization. The need for high resolution is common in computer vision applications for better performance in pattern recognition and analysis of images. High resolution is of importance in medical imaging for diagnosis.

Project Implementation Method

The methodology includes compression algorithms, some image metrics and the Convolutional Neural Network. The images collected are in the pairs of low and high resolution images. Next, they have been divided into training set and testing set with 80% to 20% ratio respectively, which are specified in the CNN model.

First and foremost step is to build the CNN. CNN consists of many layers through which we get the output. Using keras, a famous python library in order to build the CNN in python. A Sequential CNN model is implemented for the system.

Then, after defining the CNN, it needs to be compiled. We specify our training set, test set and no of epochs, in order to start training the model. While compiling it, 3 parameters have to be specified. Optimizing Algorithm, which in our case is the Adam Optimizing Algorithm. Next comes the loss function, which is binary_crossentropy since we are working with 2 classes, for more than 2 classes, categorical_crossentropy is used. Then, the last parameter is the metric for training, which is the accuracy. 

The image metrics that are involved in the project are MSE, PSNR and SSIM, to compare the reconstructed image with the original image. The MSE, PSNR and SSIM are commonly used criterion for quantitatively measuring the efficiency of image reconstruction.

MSE can be used as an image metric to evaluate the error by comparing the reference image and the one that is produced. It’s not quite effective as other metrics as it is computed pixel wise.

Another Image metric is a Peak signal-to-noise ratio (PSNR) is one of the most popular reconstruction quality measurement of lossy transformation (e.g., image compression, image inpainting). For image super-resolution, PSNR is defined via the maximum pixel value and the mean squared error (MSE) between images.

Then, another image metric is a structural similarity index metric (SSIM), it is proposed for measuring the structural similarity between images, based on independent comparisons in terms of luminance, contrast, and structures.

“In future, considering the advantages and disadvantages of above metrics, one is going to be chosen as the soul image metric in order to evaluate our final result.”

Benefits of the Project

Many applications require zooming of a specific area of interest in the image wherein high resolution becomes essential, e.g. surveillance, forensic and satellite imaging applications and sometimes, we see that images are often confronted with severe degradations (e.g., low-        resolution, low-contrast, and noise). This significantly limits the performance of face recognition systems. So, this SR based application will come in handy and will be useful in such cases.

Technical Details of Final Deliverable

The final deliverable will be with a proper use friendly user interface for any user. The dataset would be extended to a much more larger number of images, in order to convert any kind of low resolution image into a high resolution image. Image metrics would be assessed and considering their advantages and dis advantages, one would be selected and used for our final system. Through  our friendly user interface, the user would be allowed to input any low resolution image of any face, which will be processed by our system and through our model, it's high resolution image will be reconstructed with all the details that were missing from the low resolution image.

Final Deliverable of the Project

Software System

Core Industry

IT

Other Industries

Core Technology

Artificial Intelligence(AI)

Other Technologies

Sustainable Development Goals

Partnerships to achieve the Goal

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com