In recent years, with the development of deep neural networks towards large-scale, Parameter-Efficient Fine-Tuning (PEFT) has emerged as a key technique for adapting models to downstream tasks under resource-constrained environments by updating only a subset of parameters. However, existing PEFT methods like LoRA face a dilemma in deep networks: increasing trainable parameters to maintain performance escalates memory overhead, while reducing parameters may compromise model expressiveness. To address this, we propose DCK_MoE (Decomposed Convolutional Kernel_Mixture-of-Experts), a framework that integrates a dynamic routing-based MoE architecture with convolutional filter decomposition for expert network design. Specifically, each expert’s convolutional kernels are decomposed into meta kernels and meta kernel coefficients, where only the meta kernels—with significantly fewer parameters—are fine-tuned. This drastically reduces trainable parameters. Meanwhile, the MoE mechanism enhances model expressiveness, and by sharing meta kernel coefficients across experts, our framework avoids parameter explosion, achieving sublinear growth in parameter count. Experiments show that our method outperforms full fine-tuning and linear probing, surpassing LoRA in Top-1 accuracy with comparable parameters. Visualization further confirms that our gating network dynamically combines expert knowledge, enabling specialization. Our approach provides an efficient and scalable fine-tuning solution for memory-constrained environments.