随着数字图像处理技术的快速发展，图像篡改手段日益复杂，给图像真实性鉴定带来了巨大挑战。提出了一种基于双路径特征融合和双池化自适应压缩的图像篡改检测网络，以解决图像篡改检测中特征提取不足、关注点分散以及噪声与篡改特征区分困难等问题。首先，利用双路径特征融合从多角度提取图像特征，有效突出篡改区域与未篡改区域之间的差异；其次，提出了双池化自适应压缩机制，通过全局平均池化和最大池化结合，增强网络对全局上下文的感知能力，并根据输入数据的特性自适应地调整通道压缩比例，从而提高特征表征能力和篡改检测精度。此外，自适应注意力特征融合模块结合多种特征来源，利用注意力机制融合噪声、边缘及纹理信息。为了进一步优化卷积结构，引入了嵌入门控信号的残差反馈模块，以更精准地捕捉和强化任务相关特征，同时采用扩张卷积增加感受野，从而更好地捕捉上下文信息。实验结果表明，与多视角多尺度监督网络（multi-view multi-scale supervised networks, MVSS-Net）等方法相比，提出的模型在中国科学院自动化研究所CASIA数据集和哥伦比亚大学COLUMB数据集上分别将F₁得分提高了2.8%和3.3%。该方法在处理细微篡改特征时表现出较高的准确性和鲁棒性，在复杂图像篡改场景下具有良好的适应性和泛化能力。

With the rapid development of digital image processing technology, image tampering methods have become increasingly complex, posing significant challenges to the identification of image authenticity. A network for image tampering detection based on dual-path feature fusion and dual-pooling adaptive compression was proposed to address issues such as insufficient feature extraction, scattered focus, and difficulty in distinguishing between noise and tampering features in image tampering detection. Firstly, dual-path feature fusion was utilized to extract image features from multiple perspectives, effectively highlighting the differences between tampered and untampered regions. Secondly, a dual-pooling adaptive compression mechanism was proposed, through the combination of global average pooling and maximum pooling, the network's ability to perceive global context was enhanced, and the channel compression ratio was adaptively adjusted according to the characteristics of input data, thereby improving feature representation capability and tampering detection accuracy. In addition, an adaptive attention feature fusion module was used to integrate multiple feature sources, fusing noise, edge, and texture information by means of an attention mechanism. To further optimize the convolutional structure, a residual feedback module embedded with gating signals was introduced to more accurately capture and strengthen task-related features, and dilated convolution was adopted to increase the receptive field, thus better capturing contextual information. Experimental results show that compared with methods such as multi-view multi-scale supervised networks (MVSS-Net), the proposed model increases the F₁ score by 2.8% and 3.3% on the CASIA dataset from the Institute of Automation, Chinese Academy of Sciences, and the COLUMB dataset from Columbia University, respectively. The method exhibits high accuracy and robustness in handling subtle tampering features, and has good adaptability and generalization ability in complex image tampering scenarios.