International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 10 Issue: 05 | May 2023
www.irjet.net
p-ISSN: 2395-0072
CenterAttentionFaceNet: A improved network with the CBAM attention mechanism Thu Hien Nguyen1#, Danh Vu Nguyen1#, Trang Phung2* Thai Nguyen University of Education, Thai Nguyen, Vietnam Thai Nguyen University, Thai Nguyen, Vietnam ---------------------------------------------------------------------***--------------------------------------------------------------------1
2
Abstract - Convolutional Neural Network (CNN), one of common the Deep Learning models, is becoming more and more advanced, becoming the most widely used solution for most computer vision-related applications including facial recognition. Due to their high accuracy and practicality, facial recognition models play a key role in most real-world scenarios. However, the training process of these models is time-consuming and expensive. Therefore, designing a lightweight model with low computational cost and memory requirements is one of the most practical solutions for face recognition. CenterFaceNet is one of the popular lightweight networks to address the facial detection problem. In this paper, we proposed the combination of CenterFaceNet and attention modules to enhance performance while keeping the simplicity of lightweight architecture. Specifically, we propose to utilize CBAM attention that includes the Channel Attention Module and Spatial Attention Module after each block of the CenterFaceNet backbone. The test results of our proposed model on the WIDER FACE dataset show superiority to the original CenterFace model and state-ofthe-art methods.
lightweight but extremely powerful. CenterFace's network architecture is depicted in Figure 1.
Figure 1: Overall of the CenterFaceNet architecture
2. RELATED WORK 2.1. Previous Methods Continuing with related work, there are also several studies on learning multitasking [8, 9] for face detection. This approach involves the use of multiple monitoring labels to improve the accuracy of each task by using correlations between tasks. Detecting and aligning faces in a simultaneous model is widely used because the task of aligning and rearranging facial landmarks is done using the process of key feature extraction (backbone) [10] of a neural network, providing better features for the face classification task with information from face points. Similarly, RCNN significantly improved the detection performance by adding a branch to predict the faces of the subjects. State-of-the-art studies on face detection were performed using cascaded CNN methods [6], using anchor points [1113] and multitasking learning. Although each method has its own advantages and disadvantages, recent advances have shown that the anchor-based methods and its phases [14, 15] have made significant advances in both accuracy and efficiency. These methods densely sample face locations and scales on feature maps and use natural anchor points [11] or single points representing faces for regression, thus simple Simplify the training process and significantly reduce the training time. Furthermore, there have been studies on the use of attentional mechanisms to enhance detection and
Keywords: Face Detection, Attention, Deep Learning, MobileNet, CBAM.
1.INTRODUCTION Face detection is an important area of object detection in computer vision [1], which has wide applications in areas such as security [2], recognition [3], image processing [4], video classification[5], etc. The goal of this process is to find and locate faces in an image. In recent years, the development of face detection algorithms has made significant progress, thanks to the development of deep learning models and the development of neural networks e.g., CNN - Convolutional Neural Networks) [6]. The previous face detection methods have inherited the model based on the common object detection framework. The results have shown that the combination with deep learning has significantly increased the performance and accuracy of the model. However, the problem of face located prediction seems inaccuracy due to many possible results in an image. In addition, high inference time cost and large model is also very challenging. In this paper, we export a simple and effective face detection and alignment model architecture based on CenterFace [7], which is
© 2023, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 1