Abstract: Traditional digital watermarking methods are robust to attacks such as crop, noise, and deformation, but it is difficult to resist the loss of watermarking caused by image compression coding and mobile camera flipping in real world. To enhance the robustness, an end-to-end digital blind watermarking method based on frequency domain convolution is proposed by exploiting the masking property of some frequency bands in the discrete cosine transform spectrum to human eyes and the learning ability of convolutional neural network to some invisible perturbations. First, the convolutional neural network is used to build an encoding network to embed the watermarking information into the image frequency domain; second, a decoding network symmetric to the encoding network is built to extract the watermarking information from the image frequency domain; finally, the encoding and decoding networks are jointly trained and the image quality and the watermarking extraction effect are supervised. The experimental results on the MIRFLICKR show that PSNR, SSIM, LPIPS, BPP and the accuracy under-display flipping of the method reach 36.29dB, 0.951, 3.11×10^-3, 2.44×10^-3 and 93.1% respectively, which have advantages over other benchmark methods and prove the effectiveness of the method.