We propose a model based on the Lightweight Attention Residual Network (LAResNet) that integrates the advantages of attention mechanisms and residual networks. By assigning higher weights to important frequency bands and spatial features, the model enhances the accuracy of emotion recognition. Initially, multi-band differential entropy features of EEG signals are extracted and transformed into a three-dimensional feature matrix. Subsequently, the LAResNet model processes these features, focusing on the most relevant information for emotion classification tasks. Experimental results demonstrate that on the DEAP dataset, LAResNet achieves accuracy rates of 95.75%, 96.32%, and 96.11% for valence, arousal, and dominance dimensions, respectively. Compared to existing mainstream models for affective recognition, LAResNet has achieved a significant improvement in emotion recognition performance.