Facial expression recognition is a key task in the field of affective computing and human-computer interaction. Despite significant advances in this problem, benefiting from the development of Convolutional Neural Networks (CNNs) and other network architectures, the generalization ability and robustness of existing methods are often limited. To address this, we propose an efficient facial expression recognition framework called EDER-ACNN, which uses Attention CNN-based morphological Erosion and Dilation for Enhanced Recognition preprocessing. In our framework, the face image is first subjected to a series of com- posite morphological operations based on erosion and dilation. These operations serve to selectively enhance key facial features and reduce noise, making the images more suitable for subsequent CNN processing. Then, spatial transformer is integrated as an attention mechanism, which adaptively focuses on the most salient regions of the image, thereby improving the CNN’s effectiveness. Finally, the preprocessed images are fed into a channel attention-based CNN for comprehensive feature learning and emotion classification. Experimental results on the CK+ and JAFFE datasets demonstrate that EDER-ACNN achieves recognition accuracies of 97.7% and 93.8%, respectively, outperforming traditional methods and showcasing superior robustness and generalization capabilities.