| 摘要: |
| 尽管基于深度学习的图像篡改定位模型的性能不断提高,但是网络大小和参数量的增加影响了其推理速度和实际部署。知识蒸馏和模型量化是压缩模型的有效方法。其中,MGD (Mask Generation Distillation)是一种使用掩码特征的生成式特征蒸馏算法。然而,在使用该算法对图像篡改定位经典模型DenseFCN进行量化时,不同的掩码率会导致量化模型的精度差异较大。此外,在蒸馏训练中,维持固定的掩码率难以提升模型性能,尤其是在训练后期。因此,本文在MGD算法的基础上提出了一种使用动态掩码率的方法——DMGD (Dynamic MGD)。通过增加一个动态的模块,DMGD算法将掩码率可训练化。此模块将学生特征和教师特征作为输入,在蒸馏过程中根据不同阶段给出适当的逐实例和逐通道的掩码率。实验结果表明,在Defacto、PS-boundary等数据集上,相较于MGD和未使用蒸馏方式,DMGD能够使DenseFCN低比特量化模型在F1等指标上达到最优。且在跨域数据集的实验结果也显示DMGD量化模型的F1等指标也更优秀。此外,消融实验结果表明,DMGD相较于MGD的时间代价增加也非常有限。 |
| 关键词: 深度学习 图像篡改定位 模型量化 知识蒸馏 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2025.11.08 |
| 投稿时间：2023-12-12修订日期：2024-06-11 |
| 基金项目:本课题得到国家自然科学基金项目(No.62272314,No.U23B2022,No.U22B2047),深圳市基础研究项目(No.JCYJ20250604181211016,No.SYSPG20241211174032004)资助。 |
|
| Image Tampering Localization Algorithm Based on Dynamic Masking Rate and Knowledge Distillation |
| LI Sili,LIAO Guiying,TAN Shunquan,HUANG Jiwu,LI Bin |
| Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen Key Laboratory of Media Security, Shenzhen 518060, China;Guangdong Laboratory of Machine Perception and Intelligent Computing, Faculty of Engineering, Shenzhen MSU-BIT University, Shenzhen 518116, China |
| Abstract: |
| In recent years, the demand for accurate and efficient image tampering detection systems has surged, researchers have intensified their efforts to enhance the performance of such systems while also addressing the challenges posed by their deployment in real-world scenarios. One of the primary concerns has been the computational overhead associated with deep learning-based models, particularly as they become increasingly complex to achieve higher levels of accuracy. This trade-off between model complexity and inference speed has prompted the exploration of model compression techniques, with knowledge distillation and model quantization emerging as promising solutions. The Mask Generation Distillation (MGD) algorithm, which harnesses the power of masked features to distill knowledge from a teacher model to a student model, has shown great potential in compressing image tampering detection models. However, the effectiveness of MGD can be hindered by the need to maintain a fixed mask rate throughout the distillation process. This rigidity limits the adaptability of the model and can lead to suboptimal performance, especially in scenarios where the optimal mask rate varies across different stages of training. To address this limitation, the proposed Dynamic MGD (DMGD) approach introduces a dynamic mechanism for adjusting the mask rate during the distillation process. By allowing the mask rate to be trainable and adaptive, DMGD enables the student model to learn from the teacher model more effectively, leading to improved performance in terms of accuracy and efficiency. The dynamic nature of DMGD not only enhances the model's ability to capture relevant features but also mitigates the risk of overfitting by adapting to the evolving characteristics of the training data. Experimental evaluations conducted on diverse datasets, including Defacto and PS-boundary, validate the efficacy of DMGD in improving the performance of image tampering detection models. Comparative analyses with traditional MGD-based approaches demonstrate the superiority of DMGD, particularly in scenarios where the optimal mask rate varies significantly. Furthermore, the negligible increase in time cost associated with DMGD highlights its practical viability and efficiency in real-world deployment scenarios. |
| Key words: deep learning image tampering localization model quantization knowledge distillation |
