Gastrointestinal (GI) bleeding is a life-threatening condition that should be diagnosed in a timely manner, allowing the patient to be effectively treated in the clinic. Endoscopy remains the primary approach for identifying the sources of bleeding in the GI tract, where subjective interpretation of images prevails. This can vary significantly across diagnosticians. In this study, we propose a fusion-based deep learning model that combines DenseNet121 and EfficientNetB0 to facilitate automated binary classification of endoscopic images into Normal and Lesion categories. The model leverages transfer learning to extract rich and complementary feature representations from both networks, which are subsequently integrated through a fully connected classification head. To address the challenge of limited annotated data, we implemented an extensive augmentation pipeline encompassing eighteen diverse techniques ranging from geometric and photometric transformations to advanced masking and motion-based methods, thereby increasing the dataset from 113 to 1000 images per class. The model was trained for 50 epochs using a batch size of 32 and a learning rate of 0.0001. We evaluated the performance of the models with augmented data and without augmentation, using metrics such as Accuracy, Precision, Recall, F1-score, ROC curve, confusion matrix, and XAI. The results show a significant performance gain for augmented shallow data, where the proposed model achieves a classification accuracy of 96.00%. These findings underscore the potential of model fusion and data augmentation in enhancing the robustness and diagnostic accuracy of AI-driven GI bleeding detection systems in clinical practice.