Mixed Real-Life and Anime Harmful Images Classification Using Deep Residual Neural Networks and Migration Learning

Harmful image detection and classification is a significant work for network security. Earlier studies based on skin color detection using manually extracted features are inaccurate and vulnerable to environmental change. Recently deep learning-based methods, especially the convolutional neural netw...

Full description

Saved in:
Bibliographic Details
Published in2021 6th International Conference on Image, Vision and Computing (ICIVC) pp. 154 - 160
Main Authors Li, Hao, Chen, Fan, Mou, Yi, Li, Yaqin
Format Conference Proceeding
LanguageEnglish
Published IEEE 23.07.2021
Subjects
batch normalization
Degradation
Feature extraction
harmful images
image classification
mitigation learning
Network security
Neural networks
ResNet
Semisupervised learning
Skin
Training
Online AccessGet full text

Cover

More Information
Summary:Harmful image detection and classification is a significant work for network security. Earlier studies based on skin color detection using manually extracted features are inaccurate and vulnerable to environmental change. Recently deep learning-based methods, especially the convolutional neural network (CNN) based models, have received widespread attention and show promising performance. However, as the neural network depth tends to deepen, popular CNN models for harmful image classification face two problems: the degradation problem, causing a decrease in accuracy with the increase in network layers; and the overfitting problem due to the relatively small samples of harmful images, resulting in lower test accuracy and unstable performance. In this paper, we propose a classification based on deep residual neural network (ResNet) and migration learning. ResNet copes with the degradation problem by a well-designed residual block structure, preventing performance loss in deeper networks. By introducing Batch Normalization (BN) and migration learning method to ResNet, the overfitting problem during model training and initialization phases can be mitigated. Besides, a fine-grained classification is proposed for both real-life images and 2D anime images as the latter is rapidly developing, especially in online communities for young people. Experiments show that the proposed method has better fine-grained classification accuracy. The degradation and overfitting problems are well solved. The optimized network depth for online mixed harmful image classification is also discussed.
DOI:10.1109/ICIVC52351.2021.9526941