The increasing occurrence of oral and dental problems such as oral cancer is a significant global health issue. Timely identification of oral cancer is crucial for enhancing prognosis rates. Although traditional biopsy methods may cause discomfort and involve invasiveness, there is a need for a less intrusive and more convenient option. Multiple clinical research studies have investigated the application of convolutional neural network models with pre-trained weights for the categorization of oral and dental disorders. Our study presents a new deep learning model that incorporates feature fusion to enhance the categorization of abnormal oral lesions, specifically in instances of oral cancer. This novel method enhances feature extraction capabilities by combining the InceptionResNetV2 and Xception Transfer Learning models and incorporating a CBAM attention layer to maintain current feature connections. Furthermore, we utilized the Mouth and Oral Disease (MOD) dataset and the Oral Squamous Cell Carcinoma (OSCC) Cancer dataset to assess the model's effectiveness. The categorization obtained 87% average accuracy in the MOD dataset and 94% in the OSCC dataset. This study makes a significant contribution to current efforts to create non-invasive and effective methods for detection of oral cancer early, which could improve patient results and lower death rates.