Geometry problem solving, being a longstanding quandary in the domain of artificial intelligence education, has garnered significant attention. The resolution of geometry problems necessitates the amalgamation of textual depictions and geometric diagrams to construct a formal language, coupled with an understanding of geometric theorems to generate coherent sequences of solutions and ultimately attain answers. However, deficiencies in diagram parsing and modal fusion have resulted in subpar problem-solving capabilities. To proficiently extract diagram features, this paper proposes an enhanced diagram parser, DenseNet-121. Simultaneously, it implements meticulous structural and semantic analysis of textual descriptions to circumvent nuanced discrepancies in text comprehension that may lead to divergent solutions. The efficacy of this approach in geometry problem-solving has been substantiated through empirical experimentation.