Research On Deep Learning-based Facial Image Retouching Detection Method

As the high speed development of digital technology,a growing number of image processing and editing software are extensively used in image retouching,which makes digital image tampering and facial image retouching to be a common phenomenon in advertising magazines,social media dissemination,online photo sharing,and even used in certificate photos.In recent years,with the widely using of the security face recognition software,the research for facial image retouching forensics become particularly important.In this paper,we study the method of image retouching detection based on deep convolutional neural network,including retouching detection for whole facial image and local facial image.The main works are as follows:A facial image retouching detection method based on deep convolutional neural network is proposed.Considering that the inter-class difference between the retouching image and the original image in the forensic field is far less than the inter-class difference of the image classification problem in the field of computer vision,we investigate the influence of the convolution kernels with different types and sizes on the detection performance of the image retouching detection method,and then construct a deep convolutional neural network model with eight convolutional layers to achieve facial image retouching detection.The proposed network model consists of three parts:preprocessing of face extraction,feature extraction and classification.Considering that the target regions of the facial retouching detection is the face part of the image,a face extraction algorithm based on Histograms of Oriented Gradient feature is introduced in the proposed method,and the face part is extracted from the test image as the only input of the network model.In the feature extraction part,we use a back-propagation algorithm to update automatically the weights and bias of each convolutional layers.And the classification part outputs the model loss values and the classification results.The results show the effectiveness of the proposed method is satisfactory.Comparing with the classical network structures algorithm of image classification,facial retouching detection algorithm and facial makeup detection algorithm,the proposed method has better detection performance.The rapid development of digital technology has made the retouching function of image editing software no longer limited to the overall retouching,but also includes local facial retouching,such as eye enlargement,pupil color change,lip color change and other operations.Aiming at this kind of problem,a multi-scale nonlinear convolutional neural network model with fourteen convolutional layers is proposed,which includes image preprocessing,feature extraction and classification.In the feature extraction part,we propose a multi-scale convolutional layer,which enables the network model to learn the detailed features of different retouching operations;Four nonlinear multi-layer perceptron convolutional layers are added to replace the traditional linear convolutional layers,so that the network model can effectively learn the feature representation of the image retouching operation and achieve better data ion;Combining the TanH activation function with the Rectified Linear Unit activation function,a new activation function is proposed,which not only ensures that the gradient of the network is not attenuated during the training process,but also maintains the information between the neurons without loss,and further increasing the nonlinearity of the network structure.The experimental results show that the proposed method has better detection performance than other network structures for image classification and has good robustness.