General High-Pass Convolution: A Novel Convolutional Layer for Image Manipulation Detection.

Image Manipulation Detection is different from most computer vision tasks, such as Image Classification. The latter pays more attention to image content, while the former focuses more on manipulation trace. However, plain convolutional layer tends to learn features that represent image content rather than features that represent manipulation trace, which degrades the performance of traditional CNNs on Image Manipulation Detection. Inspired by constrained convolutional layer proposed by Bayar et al., we propose General High-Pass Convolution, a new form of convolutional layer which is capable of motivating CNNs to learn manipulation trace features. General High-Pass Convolution is designed to simulate a set of learnable high-pass filters to suppress the image content, thus motivating the CNNs to learn manipulation trace features which are mainly present in the high-frequency components of the image. We conduct comprehensive experiments to evaluate the effectiveness of General High-Pass Convolution. The experimental results show that General High-Pass Convolution achieves better performance than Bayar et al.’s constrained convolutional layer, and can be combined with CNN backbone networks to improve their performance on Image Manipulation Detection, such as VGG and ResNet.