With the development of deep learning technologies, table structure recognition (TSR) has become a key task in information extraction and document understanding, especially in the automatic handling of complex table layouts. Traditional TSR methods rely heavily on rules and feature extraction, making it difficult to effectively address challenges such as cell text alignment and irregular cell structures. Additionally, tables captured by mobile devices often suffer from geometric distortions, which increase the difficulty of parsing, and existing methods tend to perform poorly in such scenarios. In this paper, we propose an innovative method called SemReFixNet to efficiently capture cross-cell semantic information and spatial features for TSR. SemReFixNet include two modules named CSATIM and CellFixNet respectively. CSATIM combines the pre-trained BERT language model with a cross-attention mechanism to solving text alignment within the same cell. CellFixNet uses spatial relationships and positional features to repair missing cells caused by distortions, which improve the adaptability to irregular table layouts. Experimental results show that SemReFixNet outperforms existing methods in table structure recognition, text alignment, and missing cell repair, validating its effectiveness in complex table recognition.