Rapid growth in the complexity and size of Convolutional Neural Networks (CNNs) poses significant challenges in computational resources and energy consumption. This paper presents a novel approach to CNN optimization through a combined pruning and compression technique designed to preserve knowledge and enhance efficiency. Unlike traditional pruning methods that introduce sparsity and require specialized hardware, our method focuses on pruning entire filters based on their importance, followed by an innovative compression strategy that merges the least informative filters. This preserves the knowledge contained within these filters while maintaining the network’s structural integrity. We propose a three-step algorithm: selecting low-information filters using entropy metrics, grouping similar filters, and merging these groups to retain critical information. Our approach significantly reduces the network size without compromising accuracy, as demonstrated using the CIFAR-10, MNIST, Fashion MNIST, and USPS datasets, as well as a modified and original VGG16 architecture. The experiments carried out have shown that our method achieves a substantial reduction in the number of parameters and Floating Point Operations (FLOPs), lowering computational costs by up to 86.82% while preserving up to 99% of the original accuracy of the model. This paper contributes to the field of deep learning by offering a scalable, hardware-agnostic solution for CNN optimization, making it highly suitable for deployment in resource-constrained environments. Future work will explore the application of this method to various architectures and datasets to further validate its efficacy and versatility.