毕业生知识库

当前，随着图像编辑和生成软件的迅速发展，使得篡改图像内容或创建新图像变得越来越简单。这些生成图像（如渲染图像和着色图像）具有很高的视觉真实感，给很多重要应用领域带来了潜在的巨大威胁，例如，司法部门需要鉴别图片不是由计算机渲染软件制作的，着色图像会导致识别/监测系统给出错误的判断等。因此，鉴别图像是否由计算机生成也引起了学术界与工业界的广泛关注。

本文研究不同的计算机生成图像的鉴别问题，包括渲染图像和着色图像等，即区分图像是由照相机拍摄，还是由计算机生成。本文的主要目标是，设计高效的检测器，不仅有很高的分类精度，也有很好的泛化能力。本文考虑数据集构造、网络结构、训练方法、可视化以及理解。主要贡献总结如下：

1.提出了一种基于负样本插入的着色图像检测方法

针对当前基于手工特征或者卷积神经网络（convolutional neural network，CNN）的检测器在具有挑战性的盲检测场景中泛化能力不足，本文提出了基于负样本插入的增强训练方法来提升检测器的泛化性能。具体地，通过成对的自然图像和着色图像的线性插值自动地构造负样本，然后迭代地将这些负样本加入到原始的训练数据集来继续训练网络。

2.提出了着色图像检测的一种泛化方法

为了提高CNN取证方法的可信赖性，与此相关的一些问题需要思考和研究，例如，由CNN自动提取的判别特征的合适性以及在测试阶段对``未知''数据的泛化性能。本文以着色图像检测为案例对这些问题进行研究，并得到一些有用的启示。另外，基于CNN第一层对取证性能影响的定量分析，本文提出了着色图像检测的一种泛化方法，通过组合具有不同设置的第一层的CNN模型的决策结果，来提高着色图像检测的泛化性能。

3.提出了一种基于CNN的自然图像与计算机渲染图像的鉴别方法

针对基于手工特征方法的识别能力有限，本文引入了一个基于CNN的通用框架。本文方法在由多源数据组成的具有挑战性的公开数据集上表现出良好的性能，并且对几种典型的后处理操作具有很好的鲁棒性。另外，本文首次尝试使用先进的可视化工具来理解CNN学习到的关于自然图像和计算机渲染图像差异的知识。

4.提出了一种基于特征多样性增强和对抗样本的计算机渲染图像鉴别方法

针对当前CNN检测器的分类能力有限，尤其是泛化能力不足，从网络结构和网络训练两个方面进行了探索。为了进行实验研究，本文收集了4个高品质的渲染数据集。网络结构方面，设计一个新的两支路CNN网络，该网络的两个支路的第一层使用不同的初始化方法，来丰富深度特征的多样性；网络训练方面，提出一种新的以模型为中心的方法来生成负样本，继而采用增强训练来进一步提升CNN检测器的泛化能力。

Nowadays, with the tremendous advances of image editing and generating software, it has become easier to tamper with the content of images or create new images. These generated images, such as computer rendered (CR) image and colorized image (CI), have high-quality visual realism, and potentially throw huge threats to many important applications, for examples, the judicial departments need to identify that the pictures are not produced by computer rendering software, colorized images can cause recognition/monitoring systems to give incorrect judgments, and so on. Therefore, the identification of whether an image is generated by the computer has also attracted widespread attention in academic and industrial communities.

This thesis studies the identification of different computer-generated images including CR image and CI, namely, identifying whether an image is acquired by a camera or generated from the computer. The main objective is to design an efficient detector, which has high classification accuracy and good generalization capability. This thesis considers the dataset construction, network architecture, training methodology, visualization and understanding. The main contributions are summarized below:

1. A colorized image detection method based on negative sample insertion is proposed

Considering the insufficient generalization ability of current detectors based on hand-crafted features or convolutional neural network (CNN) in the challenging blind detection scenario, an enhanced training method based on negative sample insertion is proposed to improve the generalization of the detector. Specifically, negative samples are automatically constructed through linear interpolation of paired natural images and colorized images, and then are iteratively added into the training set to continue to train the model.

2. A generalization method for colorized image detection is proposed

To enhance the trustworthiness of CNN-based forensic methods, some problems regarding this need to study and answer, including for example the appropriateness of the discrimination information automatically extracted by CNN, the generalization performance on ``unseen'' data during the testing phase, etc. This thesis takes the colorized image detection as an example to study these problems and gets some useful hints. Moreover, based on the quantitative analysis of CNN's first layer on forensic performance, this thesis proposes a generalization method for colorized image detection, specifically, combining decision results from CNN models with different settings at the first layer of the network to improve the generalization performance of colorized image detection.

3. A method for the identification of natural image (NI) and CR image based on CNN is proposed

Considering the limited recognition ability of hand-crafted-feature-based methods, a generic framework based on CNN is introduced. This method shows good performance in the challenging public datasets comprising images of heterogeneous origins, and also demonstrates strong robustness against several typical post-processing operations. In addition, this thesis first attempts to understand what the CNN has learned about the differences between NI and CR image by using advanced visualization tools.

4. A CR image identification method based on the enhancement of feature diversity and adversarial sample is proposed

Concerning the limited classification capability of current CNN detectors, especially the generalization performance, this thesis explores from two aspects of network structure and network training. To study this challenging problem, this thesis collects four high-quality CR image datasets. For the network structure, a new two-branch CNN is designed. The first layer of the two branches of the network uses different initialization methods to enrich the diversity of deep features. For the network training, a novel model-centric method is proposed to generate negative samples, and then applying the enhanced training to further improve the generalization of the CNN detector.