Document Image Translation (DIT) requires accurate preservation of both textual semantics and spatial layout during cross-lingual conversion. Existing approaches predominantly rely on geometric heuristics for post-OCR text clustering, often failing to capture the semantic coherence essential for high-quality translation. To address this, we propose a novel Layout-aware Semantic Paragraph Clustering algorithm designed to reconstruct coherent paragraphs from fragmented OCR results. Our method operates through an iterative framework composed of three synergistic modules: (1) a Spatial Neighbor Selection module that identifies spatially proximate OCR fragments based on geometric constraints, (2) a Semantic Concatenation Model that evaluates semantic coherence for candidate fragment merging, and (3) a Completeness Judgment Model that determines whether aggregated segments constitute semantically independent paragraphs. Through iterative optimization, our framework reconstructs semantically coherent and spatially consistent paragraph structures, significantly enhancing downstream translation quality. Experiments on DIT700K and cross-domain evaluations demonstrate substantial improvements over SOTA methods, with BLEU score improvements of up to 17.37 points on complex layouts. Our framework serves as an effective post-processing component that enhances both traditional cascaded systems and modern multimodal language models.