Aiming at the insufficient performance of existing deep learning based digital image robust watermarking schemes under attacks such as JPEG compression and cropping, this paper proposes a digital image robust watermarking model with multi-scale attention (DIRW-MSA). The model adopts an “encoder-noise layer-decoder” structure. The encoder realizes feature extraction and cross-group fusion through the designed multi-scale channel attention fusion (MSCAF) module; meanwhile, a spatial attention mask (SAM) layer is constructed to guide the watermark to adaptively adjust the embedding strength according to image features; the decoder uses MSCAF to extract multi-scale features to improve the accuracy of watermark extraction. Experimental results show that under JPEG attack training with different quality factors, both the peak signal-to-noise ratio (PSNR) of the encoded images and the bit error rate (BER) of watermark extraction of the proposed algorithm are superior to those of existing comparison algorithms; under combined noise attack training, the average PSNR of the encoded images generated by the proposed model reaches 38.6738dB, and the BER of watermark extraction is all lower than 0.6%, which is better than the existing comparison algorithms.