Network pruning is a powerful and effective technique to reduce model complexity while maintaining high performance. Unstructured pruning has been commonly used in model compression due to its ability to achieve high compression ratios by eliminating a large number of parameters. However, unstructured pruning may pose challenges for hardware compatibility due to irregular memory access patterns and inefficient storage allocation. In this paper, we introduce a novel pruning method called SCP (Structure Combination Pruning) that addresses these challenges. SCP ensures compatibility with hardware constraints while still achieving a high compression ratio. Specifically, SCP merges complementary structures following group-wise and stripe-wise unstructured pruning, thereby achieving pruning goals while maintaining hardware compatibility. Additionally, SCP considers the impact of different channels numbers in the convolutional layer on the inference time of different target devices. Our extensive experiments demonstrate that SCP outperforms both structured and unstructured pruning methods. By combining the benefits of structured and unstructured pruning while mitigating their limitations, SCP represents a significant advancement in model compression techniques. Code is available at this url .