Video captioning aims to translate visual content into natural language descriptions. Although existing methods have achieved notable progress, they often focus on the overall video content while overlooking abundant semantic and detailed information within frames, consequently struggling to generate accurate and comprehensive descriptions. To address this limitation, we propose the Semantic Emphasis and Detail Supplementation (SEDS) framework, designed to concurrently model frame-level key semantics and inter-frame supplementary details for enhanced caption generation. The SEDS architecture integrates two core modules: 1) The Key Semantic Extraction (KSE) Module executes a focused semantic reconstruction task to guide attention toward core frame-level semantics, capturing more accurate and richer video semantics; 2) The Differential Detail Perception (DDP) Module employs frame-instance contrastive learning to direct attention to inter-frame variations, thereby capturing easily overlooked details that enhance descriptive completeness. Comprehensive experiments on MSVD and MSR-VTT datasets demonstrate the effectiveness of our approach, achieving BLEU-4 scores of 61.9 and 47.7, and CIDEr scores of 112.7 and 57.7, respectively. Our code is available at https://github.com/smile0063/SEDS .