| 摘要: |
| 针对当前Web不良图像内容监管和智能审核需求快速增加，人工和传统算法审核监管在规模、灵活性和响应时间上存在的不足，以及现有相关暴力图像数据样本的缺乏，提出一种基于深度模型迁移学习的Web图像内容审核方法。首先，收集自建暴力图像样本数据集，并对其进行数据增强和图像增强处理；其次，选择ImageNet数据集上训练完成的VGG16(Visual Geometry Group)和Resnet50(Residual Neural Network)两种典型的预训练深度神经网络模型进行迁移学习；最后，通过共享通用视觉特征、模型权重参数迁移以及微调，最终优化得到图像内容审核模型；通过对比验证实验研究，发现图像内容审核模型识别性能明显优于现有其他方法，正确率达到了95%以上，能满足实际应用需求。 |
| 关键词: 特征提取 卷积神经网络 残差网络 迁移学习 图像审核 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Web Violence Image Recognition Based on Transfer Learning |
|
FENG Yu-ting，TENG Xian-feng，GUO Yu-tang1,2
|
|
1.School of Computer Science and Technology， Hefei Normal University，Hefei 230601， China;2. School of Computer Science and Technology， University of Science and Technology of China， Hefei 230026， China
|
| Abstract: |
| With the popularity of smart devices， the rapid development of Internet， social and multimedia software technologies， the accessibility and self-control of images have been greatly improved， and the demand for pornographic， violent and other undesirable image content supervision and intelligent review has rapidly increased. In view of the shortcomings of manual and traditional algorithm image content review and identification in scale， flexibility and response time， and the lack of existing related image data samples， a web image content review model based on deep transfer learning is proposed. According to the self-built violent image sample data set， two typical pre-trained deep neural network models， VGG16 and Resnet50， were selected for transfer learning， and the image content review model was finally optimized through fine-tuning of the weight parameters. Comparative experiments show that the recognition performance of the image content review model is significantly better than other existing methods， and the accuracy rate is more than 95%， which can meet the actual application needs. |
| Key words: feature extraction convolutional neural network residual network transfer learning image review |
