基于多尺度前馈融合结构的重采样因子估计算法

AEstimationAlgorithm of Resampling Factor Based on Multi-Level Feedforward Fusion Structure

【目的】重采样是掩盖图像篡改痕迹的重要手段,为了更加精确地实现对重采样缩放参数的检测,验证图像信息的真实性,本文提出一种基于多尺度前馈融合结构的重采样因子估计算法。【方法】在预处理层中,首先使用两个线性高通滤波器得到重采样图像的残差特征,抑制图像内容带来的影响,放大区域内像素之间的关联性,其次利用4个低阶高通滤波器在不同方向上强化像素的梯度特征,该算法的主体结构为卷积神经网络,在网络的不同层级处提取出多尺度重采样分类痕迹,结合注意力机制,形成多尺度残差融合模块(Multiscale Residual Fusion Module, MRFM),补偿卷积过程中重采样信息的丢失,标定特征信息传递过程中的有效性,同时去除信息冗余,加速网络收敛。【结果】实验表明,本文所提算法的网络增益由预处理层和多尺度残差融合模块共同决定,准确性明显高于对比的其他算法,尤其在强噪声的干扰下,本文所提算法具有明显的优势。

[Objective]Resampling is an important measure to cover the traces of image tampering.In order to accurately detect resampling scaling parameters and verify the authenticity of image information,this paper proposes a resampling factor estimation algorithm based on the multi-scale feed-forward fusion structure.[Methods]In the preprocessing layer,the residual charac-teristics of the resampled image are obtained by using two linear high-pass filters,the impact of image content is suppressed,the correlation between pixels in the region is enlarged,and then the gradient characteristics of pixels are strengthened in different directions by using four low-order high-pass filters.The main structure of the algorithm is a convolutional neural network,and multi-scale resampling classification traces are extracted at different levels of the network,combined with the attention mechanism.The Multiscale Residual Fusion Module(MRFM)is formed to compen-sate for the loss of resampled information during convolution,achieve effective transmission of the calibration characteristic information,and remove information redundancy to accelerate network convergence.[Results]Ex-periments show that the network gain of the proposed algorithm is determined by the pretreatment layer and the multi-scale residual fusion module,and the accuracy is significantly higher than that of other algorithms in com-parison,especially under the condition of strong noise interference of strong.The proposed algorithm is of obvi-ous advantages.