The inherent challenge in computer vision lies in efficiently modeling both fine-grained spatial details and global semantic context. Contemporary architectures often achieve this through computational redundancy, limiting their deployment on resource-constrained devices. We present PyraFuseNet (Dual-Path Pyramid Fusion Network), introducing three key innovations: (1) parallel specialized pathways for local and contextual feature extraction with minimal overlap, (2) dynamic pyramid fusion using learned channel-wise attention for optimal multi-scale feature combination, and (3) complexity-aware training reducing FLOPs by 55.56% through gradient-guided feature pruning. Extensive evaluation demonstrates state-of-the-art performance across MNIST (99.72%), Fashion-MNIST (96.02%), KMNIST (98.71%), CIFAR-10 (92.22%), and CIFAR-100 (85.26%), surpassing ResNet-18 while using 55% fewer computations, making it a viable candidate for resource-efficient AI systems. The proposed architecture achieves this efficiency while maintaining competitive parameter efficiency, demonstrating a transformative approach to resource-constrained deep learning.