基于深度学习的图像局部模糊识别

Local Blur Detection of Digital Images Based on Deep Learning

数字图像中常用模糊操作隐藏或抹去篡改的痕迹.为此,针对常用的高斯模糊、均值模糊及中值模糊操作的识别问题,构建了一种卷积神经网络模型,并给出其网络拓扑结构.在传统的卷积神经网络模型中添加一个信息处理层,提取出输入图像块的滤波频域残差特征,以提高网络模型对一次滤波与二次滤波操作的识别性.实验结果表明,所提方法的准确率较以往传统方法有较大提升,且泛化性能优越,能检测出主流的线性和非线性滤波操作。

Blurring is generally a post-operation to conceal or remove the trace of tam- pering. In this paper, a new convolutional neural network model is proposed, and the corresponding network topology is presented to handle the problems in the detection of blur operations, such as Gaussian blur, average blur, or median blur. An information pro- cess layer is added into the conventional convolutional neural network to extract the residual features of filtering frequency domain, accordingly, improving the accuracy of blur detec- tion between the first-order and the second-order filtering operations. Experimental results demonstrate that the proposed method performs a higher accuracy in blur detection than traditional methods, and is able to discriminate between the common linear and nonlinear blur operations.