PDF(Pubmed)
Abstract：
Deep learning-based methods have been proven excellent performance in detecting pornographic images/videos flooded on social media. However, in a dearth of huge yet well-labeled datasets, these methods may suffer from under/overfitting problems and may exhibit unstable output responses in the classification process. To deal with the issue we have suggested an automatic pornographic image detection method by utilizing transfer learning (TL) and feature fusion. The novelty of our proposed work is TL based feature fusion process (FFP) which enables the removal of hyper-parameter tuning, improves model performance, and lowers the computational burden of the desired model. FFP fuses low-level and mid-level features of the outperforming pre-trained models followed by transferring the learned knowledge to control the classification process. Key contributions of our proposed method are i) generation of a well-labeled obscene image dataset GGOI via Pix-2-Pix GAN architecture for the training of deep learning models ii) modification of model architectures by integrating batch normalization and mixed pooling strategy to obtain training stability (iii) selection of outperforming models to be integrated with the FFP by performing end-to-end detection of obscene images and iv) design of TL based obscene image detection method by retraining the last layer of the fused model. Extensive experimental analyses are performed on benchmark datasets i.e., NPDI, Pornography 2k, and generated GGOI dataset. The proposed TL model with fused MobileNet V2 + DenseNet169 network performs as the state-of-the-art model compared to existing methods and provides average classification accuracy, sensitivity, and F1 score of 98.50%, 98.46% and 98.49% respectively.
摘要：
基于深度学习的方法已被证明在检测社交媒体上充斥的色情图像/视频方面具有出色的性能。然而,在缺乏大量但标记良好的数据集的情况下，这些方法可能遭受欠/过拟合问题，并且可能在分类过程中表现出不稳定的输出响应。为了解决这个问题，我们提出了一种利用迁移学习（TL）和特征融合的自动色情图像检测方法。我们提出的工作的新颖性是基于TL的特征融合过程（FFP），它可以消除超参数调整，提高模型性能，并降低所需模型的计算负担。FFP融合性能优于预训练模型的低级和中级特征，然后转移学习的知识以控制分类过程。我们提出的方法的主要贡献是：i）通过Pix-2-PixGAN架构生成标记良好的淫秽图像数据集GGOI，用于训练深度学习模型ii）通过集成批量归一化和混合池化策略来修改模型架构以获得训练稳定性（iii）通过对淫秽图像进行端到端检测来选择性能优于FFP的模型，以及iv）设计基于TL的淫秽图像检测方法的最后一个融合层。对基准数据集进行了广泛的实验分析，即NPDI,色情2k,并生成GGOI数据集。与现有方法相比，所提出的具有融合MobileNetV2DenseNet169网络的TL模型作为最先进的模型执行，并提供平均分类精度，灵敏度,F1得分为98.50%，分别为98.46%和98.49%。
