Image splicing tamper detection based on two-channel dilated convolution
Yan Li, Cheng‐Xiang Wang · 2022
Image splicing tampering detection can provide technical support for the identification of authenticity and integrity of images in various fields by using specific methods to quickly detect whether tampering has occurred in images. Most of the existing splicing tamper detection methods pay more attention to the detection accuracy of tamper image, ignore the importance of tamper region location, and can only provide rough tamper region location map. To solve the above problems, we proposed an image splicing tamper detection model based on dual-channel dilated convolution, which combined the deep features and shallow features of the image to accurately locate the tamper area. In this model, the first channel extracts noise features through a set of high-pass filters to find noise inconsistencies between the real area and the tampered area. The second channel extracts RBG image features through dilated convolution and locates the tampered area in combination with the attention mechanism. Then, for each region of interest, the features generated by the two channels are fused together through bilinear pooling layer. Finally, tamper classification and boundary box regression are performed. We performed experimental analysis on NIST16 and CASIA datasets, and compared with the other five models, this model has better positioning performance, with the F1 index increased by 0.026 and 0.051, respectively, compared with other advanced methods.