The effectiveness of digital media advertisements increasingly relies on the accurate identification and analysis of user emotions. In this context, an improved ResNet-50 model is proposed to more accurately analyze users’ emotional responses to advertising content. The method optimizes the traditional ResNet-50 architecture by introducing an attention mechanism as well as a modified Sigmoid cross-entropy loss function, which improves the accuracy of emotion recognition. A regional convolutional neural network as well as a bi-directional encoder representation of the transformer are introduced to compare with the research method. The test results show that the F1 value of the improved model reaches 0.946, which is 0.084–0.101 higher than the comparison model; its Recall value reaches 0.938, which is 0.044–0.078 higher than the comparison model. Meanwhile, after a series of evaluations by a series of industry experts, the method receives the highest evaluation in the field of advertisement design, which not only confirms its effectiveness in practical applications, but also foretells its potential application value in ad content customization, user experience optimization, and personalized ad recommendation systems. These results indicate the innovative application of sentiment analysis using deep learning in digital advertising design, and provide strong technical support for future research and practice in related fields.