The work aims to apply the deep learning to the digital watermarking and propose a digital watermarking algorithm combining spatial domain and frequency domain based on multi-scale expanded convolutional attention (SF-ACA), so as to improve the invisibility and robustness of images while concealing information. The network framework of this algorithm consisted of two parts:a generator composed of ACA and SFE and a decoder. Among them, the MCA module in the ACA network combined three dilation convolutions with varying atrous rates for feature extraction of carrier images with multi-scale fusion, so that the carrier images could conceal the watermark information more effectively. The SFE combined fast Fourier convolution blocks to capture complementary information in the spatial and frequency domains with varied widths of perceptual fields to collect the feature information of the watermark more effectively and enhance the invisibility of the secret information and robustness. According to experimental findings, the PSNR value of the proposed watermarking method was 38.81 dB which was improved by 7.78% in comparison to the UDH method while concealing a color image of equal size to the carrier image. The watermarked image had a hiding capacity of 4 096 bits, and the method improved the extraction accuracy under Dropout, Gaussian noise, and JPEG attacks by 5.38%, 10.5%, and 1.65%, respectively, meeting the requirement of invisibility and achieving strong robustness. When the hiding capacity is high, the method described in this study performs better in terms of robustness and invisibility.