Research On Vehicle License Image Tampering Detection Based On Region-Based Convolutional Neural Networks

As a crucial legal document,a vehicle license’s authenticity is essential for maintaining the traffic system and social order.However,with the increasing number of vehicles,the authenticity and integrity of vehicle licenses are constantly threatened.Traditional image tampering detection methods rely on prior knowledge of the image and have low accuracy,while machine learning algorithms can independently extract effective features from data,which is helpful for improving the accuracy of image tampering detection algorithms.In this thesis,the R-CNN object detection algorithm is used to model the problem of detecting tampered areas in images.The main research work is as follows:(1)A license image splicing simulation method based on segmentation network models is proposed.A sufficient amount of tampered license images is needed as a dataset in license image tampering detection research,but it is difficult to obtain such datasets in reality.By analyzing existing semantic segmentation and instance segmentation networks,and considering the excellent performance of U-NET on small training sets,the high accuracy of Mask R-CNN object detection,and the uniqueness of instance segmentation in driving license images,this study uses the aforementioned two networks to generate tampered samples for subsequent tampering detection model training.(2)A multi-source feature region convolutional neural network image tampering detection method is proposed.To improve the low accuracy and detection failure caused by hidden tampering traces in tampering detection tasks,this method proposes inputting multi-source features into the target detection network.Multiple convolutional layers are added to the Cascade R-CNN network,and non-homogeneous image features are input and fused into the subsequent detection head for classification to improve detection accuracy.In order to generate various feature distribution maps of the original image,multiple filters are applied to the original image before inputting it into the convolutional neural network.The model is trained to achieve optimal performance and ultimately achieve effective detection of suspicious areas in the image.(3)Multi-source feature fusion algorithms are introduced into the region convolutional neural network.In the multi-source feature region convolutional neural network,features from multiple sources need to be fused before being classified by the subsequent classifier.In this thesis,multiple convolutional neural networks are used to obtain non-homogeneous feature distribution maps,and then bilinear pooling strategy is used to fuse features pairwise,and finally all fused features are connected.The multi-source feature fusion strategy effectively applies the steganographic features of the license image to the classification subtask of the region proposal network and achieves good results.In summary,this thesis models image tampering detection as an object detection problem,improves the basic Cascade R-CNN network to obtain a new network structure for detecting image tampering,effectively improving model accuracy.Compared with existing detection algorithms,the accuracy,recall,and F1-score of the method in this thesis are better than existing methods,proving the effectiveness of the proposed method.