VCKNet: An adaptive, modular, and lightweight Variable-sized Convolutional Kernel Network

We introduce VCKNet (Variable-sized Convolutional Kernel Network), an adaptive, modular, and lightweight Convolutional Neural Network (CNN) with three kernel-scale branches, channel attention, and dynamic fusion. These components recalibrate scale-aware features to reduce redundancy and support discriminative learning. The proposed architectural design preserves computational efficiency while maintaining a clear structural organization. Experiments on standard benchmarks and in-the-wild datasets show that VCKNet effectively captures multiscale structure, achieving performance competitive with that of deeper state-of-the-art models. Statistical analysis across repeated runs further confirms VCKNet’s stability. Computational cost and deployment efficiency analyses further support its suitability for resource-constrained and production-oriented scenarios where flexibility, stability, and conceptual clarity are essential.