<p>Convolutional neural networks (CNNs) have revolutionized computer vision, yet their high computational cost remains a barrier to deployment on resource-constrained platforms such as edge devices and mobile systems. In modern AI pipelines, CNN optimization is typically performed offline on high-performance computing (HPC) infrastructures, while the resulting compact models are deployed across edge and cloud–edge environments. In this paper, we propose an efficient CNN compression strategy based on filter pruning, designed for scalable inference without sacrificing accuracy. We introduce a novel correlation circle filter pruning method that identifies and removes redundant filters using principal component analysis (PCA)-based correlation analysis. This is embedded into a broader multi-criteria filter pruning framework that combines correlation and norm-based criteria for more informed pruning decisions. To address the limitations of one-shot pruning under high compression, we further propose an iterative pruning strategy with intermediate fine-tuning, enabling the network to progressively adapt and recover accuracy. The resulting pruning workflow is computationally intensive but highly parallelizable, making it well suited for execution on multi-GPU and HPC platforms. Experimental results on CIFAR-10 and ImageNet using ResNet architectures show substantial reductions in computational cost, achieving up to 54% FLOPs reduction on ResNet-56 with improved accuracy and 49% reduction on ResNet-50 with only a 0.10% top-1 accuracy drop and no top-5 loss. Beyond image classification, the proposed framework is further validated on a large-scale object detection task using a different CNN architecture (EfficientDet-D0) on the MS-COCO dataset, demonstrating its applicability to computationally intensive and deployment-oriented computer vision pipelines.</p>