Food recognition is a challenging problem in comparison with conventional image recognition due to the significant resemblances between various food categories. Hence, precise categorization of food images has grown progressively vital in diverse applications. This paper introduces an innovative method for quickly categorizing images of fast food. The method leverages transfer learning and also incorporates a tailored attention mechanism. Our method is built upon the base DenseNet169 architecture, and then fine-tuned for the specific task of fast food classification thereby, achieving an accuracy of 96.26%. We employ a two-stage transfer learning procedure: initially, we leverage the pre-existing weights of DenseNet169 trained on ImageNet, followed by fine-tuning the model on a comprehensive food image dataset. Finally, we adapt the model to our specific fast food dataset. This methodology enables the model to acquire general characteristics connected to food prior to focusing on specific attributes of fast food. In addition, we incorporate a tailored attention mechanism that allows the neural network to concentrate on the most distinguishing areas of food images, hence improving the accuracy of categorization. The dataset is imported from Kaggle repository and comprises of ten labels of fast food images. The experimental results obtained outperforms both the conventional machine-learning techniques and the standard deep learning models in the classification of fast food with an accuracy of 96.26%. The novelty of this paper lies in its two-stage transfer learning approach combined with a customized attention mechanism specifically designed for fast food categorization. And, this work has not been addressed so far for fast food classification to the best of our knowledge.