Knowledge graph (KG) leverage rich semantic relationships to provide effective auxiliary information for recommendation. Recently, KG-enhanced recommendation based on graph neural networks (GNN) has gained prominence as a research focus. However, GNN-based methods suffer from inherent limitations, namely sparse interaction signals and KG noise, which cause item representations to deviate from their true features and impede accurate modeling of user preferences. To address these challenges, we propose the Knowledge Enhanced Global Graph Contrastive Denoising method for Recommendation (KGCD). Specifically, KGCD employs a neighbor-aware smoothing mechanism to extract semantic features of items from the KG and integrates them into user-item interaction. A path-aware GNN is then applied to learn global representations. To tackle interaction sparsity in the KG, we introduce a knowledge enhancement strategy that aggregates higher-order relational information to enrich the neighborhood representations of nodes. Furthermore, contrastive learning is utilized to effectively denoise the knowledge information.