In this paper, we present a new image forgery detection method via a mask filter banks which is consisted with the designed mask filters to extract the features of different channels of image and a modified a ResNet to classify the input image is tempered or not. The proposed model is proved to be capable for copy-move forgery and splicing image detection. In the mask filter layer, we first convert the image from spatial domain to frequency domain, then extract the image edge information of each channel by element-wise with the designed mask matrix. Finally, edge and noise information features of different channels were fused as feature vectors fed to a trained ResNet to do classification. Experiments on three standard datasets: the copy-move forgery image datasets MICC-F220 and MICC-F2000, splicing image manipulation datasets Columbia demonstrate that proposed method get better results than the original colour image as input method and also outperform some existing works.