Research On Facial Image Enhancement Based On Attention Mechanism

Facial recognition is the most important biometric feature of an individual,and it is widely used for check-in and attendance in unit management,smart access control in urban security,and facial payment in unmanned supermarkets.In public security operations,facial recognition is often used for personnel control and investigation,cross age fugitive tracking,and the retrieval of abducted children.Most facial images obtained in industries are obtained under restrictive conditions and with the cooperation of the subject,while most facial images obtained in the process of public security law enforcement are obtained under non restrictive conditions and without the subject’s sensitivity^[1].For example,facial images captured in video surveillance may have low clarity due to factors such as device performance,surrounding environment,weather conditions,and lighting brightness.High noise and uneven brightness can cause loss of facial image content and affect the accuracy of comparison results.In recent years,deep learning methods represented by Neural Networks（NN）^[2]have brought rapid development to the field of computer image and vision,including low-quality facial image restoration.It can find the mapping relationship between low-quality facial images and the original facial image through network training,and thus perform the restoration of low-quality facial images.In the subdivision problem of low-quality facial image restoration,the restoration effect based on neural network methods surpasses traditional methods.The good restoration effect and flexible adaptability to different scenes have attracted more and more scholars’attention.This article is based on neural networks and conducts research from two directions:fusion attention mechanism and facial image enhancement,proposing new solutions to improve the quality of facial image input in monitoring systems.The main research content and contributions of this article include:（1）Propose a Cycle GAN based facial image colorization methodIn response to the issues of inaccurate color and loss of details in grayscale facial image colorization technology,this paper proposes an improvement plan,which includes embedding the self attention mechanism ACmix module into the network model Cycle GAN to weight and strengthen key facial information;Introducing the evolution operator to adaptively extract more facial information;A composite loss function is designed to calculate the error between the real image and the generated image to improve the authenticity and naturalness of the color face image.Quantitative and qualitative evaluations of the improved model were conducted on the Multi PIE and Georgia Tech Face Database datasets.Experiments showed that the proposed model can achieve more realistic and natural facial colors compared to existing methods.（2）Propose an improved facial image super-resolution reconstruction method based on mixed attention mechanismTo address the issues of traditional super-resolution reconstruction models not being able to focus on useful facial feature information structures and the appearance of artifacts in the face,based on the traditional Cycle GAN network structure,this paper proposes an improvement plan including:proposing a cyclic structure facial image enhancement algorithm;By combining channel attention mechanism,we aim to increase the network’s attention to facial features and enhance the range of input information utilized by the model;Optimize the loss function of the network,introduce the structure awareness loss function into the traditional Cycle GAN loss function,and combine it with L1 pixel loss to ensure the spatial information of the image and further enhance the performance of the model.The experimental results show that the model proposed in this article has better improvements in image quality and details compared to classical algorithms.