Convolutional neural networks (CNNs), characterized by their feed-forward and backpropagation architecture, have shown exceptional performance in various computer vision tasks. However, their effectiveness depends on careful choices of hyperparameters. These include the learning rate, batch size, dropout rate, and kernel size. Optimizing CNN hyperparameters is time-consuming and resource-intensive, relying on manual tuning. Recent advancements have introduced sophisticated optimization algorithms, like Bayesian optimization, random search, grid search, and evolutionary algorithms, automating this process and enhancing model performance. This study introduces an enhanced dynamic Salp swarm algorithm (EDSSA) for optimizing CNN hyperparameters. EDSSA addresses the challenge of variable-length individuals, enhances the search equation, and uses a hybrid good-point set (GPS)-based population initialization technique. EDSSA achieved 87% accuracy on the CIFAR-10 dataset, surpassing all earlier methods. Note that this may be the performance specifically for this dataset or for this architecture of CNNs. Though overly complicated data patterns had a few failures, the performance was particularly good in other datasets, thus showing generalization away from CIFAR-10. In practice, there are still some challenges for EDSSA, such as the high computational cost for a large-scale dataset and fine-tuning in other applications for image classification. Despite these, robust performance with fast convergence could imply deep practical implications of EDSSA because of automation for CNN hyperparameter optimization, saving time and resources.