Scholarlink

<p>The ability to create fake images, such as incorporating one’s own face with another, has been significantly made easier by sophisticated image editing software. One of the most difficult challenges is to improve multi-classification performance of an unbalanced dataset of real and fake faces. We investigated and compared 16 different face classification methods: MobileNet, MobileNetV2, NASNetMobile, ResNet50, DenseNet121, ResNet50V2, DenseNet169, ResNet101V2, ResNet152, DenseNet201, InceptionV3, ResNet152V2, Xception, InceptionResNetV2, EfficientNetV2M and our proposed model. After analyzing several face classification techniques, we developed a novel hybrid model that combined an improved version of MobileNetV2+convolutional architecture for fast feature embedding (CAFFE) named as “CMNV2" included in methodology. Goal of this study is to add 6 additional layers to pre-trained model structures previously present to increase accuracy of CMNV2 model for classification and prediction of real and fake faces of real-time and photo images. In this work, multi-classification of unbalanced dataset of 2040 images containing 4 categories: real_face, fake_face_easy, fake_face_mid and fake_face_hard is trained and tested with deep learning (DL) models of Keras applications. Our CMNV2 Proposed model, which combines deep neural networks (DNNs), transfer learning (TL) and deep learning architecture performed better than 15 standard methods in extracting image features for face classification (FC) and prediction. Through combing their feature sets at a particular layer, the CMNV2 model hybridized using MobileNetV2 and CAFFE, enabling the model to capture both generalized and specialized information. Transfer learning is then used for fine-tuning on the particular dataset. Our CMNV2 model also outperformed other 15 traditional face classification models with 9 performance metrics evaluation, achieving 98.53% percent accuracy, a 1.47% error rate with less computational time and fewer parameters. The dataset and source code used in our article are made available on github platform <a href="https://github.com/AnilKumargithu/Real-and-Fake-Faces-Classification-and-Prediction">https://github.com/AnilKumargithu/Real-and-Fake-Faces-Classification-and-Prediction</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A multi-level classification and prediction of unbalanced dataset real and fake faces using hybrid CMNV2 and deep learning methods

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • B. Anil Kumar,
  • Neeraj Kumar Misra

摘要

The ability to create fake images, such as incorporating one’s own face with another, has been significantly made easier by sophisticated image editing software. One of the most difficult challenges is to improve multi-classification performance of an unbalanced dataset of real and fake faces. We investigated and compared 16 different face classification methods: MobileNet, MobileNetV2, NASNetMobile, ResNet50, DenseNet121, ResNet50V2, DenseNet169, ResNet101V2, ResNet152, DenseNet201, InceptionV3, ResNet152V2, Xception, InceptionResNetV2, EfficientNetV2M and our proposed model. After analyzing several face classification techniques, we developed a novel hybrid model that combined an improved version of MobileNetV2+convolutional architecture for fast feature embedding (CAFFE) named as “CMNV2" included in methodology. Goal of this study is to add 6 additional layers to pre-trained model structures previously present to increase accuracy of CMNV2 model for classification and prediction of real and fake faces of real-time and photo images. In this work, multi-classification of unbalanced dataset of 2040 images containing 4 categories: real_face, fake_face_easy, fake_face_mid and fake_face_hard is trained and tested with deep learning (DL) models of Keras applications. Our CMNV2 Proposed model, which combines deep neural networks (DNNs), transfer learning (TL) and deep learning architecture performed better than 15 standard methods in extracting image features for face classification (FC) and prediction. Through combing their feature sets at a particular layer, the CMNV2 model hybridized using MobileNetV2 and CAFFE, enabling the model to capture both generalized and specialized information. Transfer learning is then used for fine-tuning on the particular dataset. Our CMNV2 model also outperformed other 15 traditional face classification models with 9 performance metrics evaluation, achieving 98.53% percent accuracy, a 1.47% error rate with less computational time and fewer parameters. The dataset and source code used in our article are made available on github platform https://github.com/AnilKumargithu/Real-and-Fake-Faces-Classification-and-Prediction.