Transfer learning is a powerful technique used to leverage knowledge learned from one domain and transfer it to another domain, reducing the time needed to train a machine learning model. This paper presents a novel approach to apply transfer learning to improve accuracy in image recognition tasks. The authors first use classical transfer learning algorithms to analyze a publicly available image dataset and extract features relevant to the recognition problem. Then, they combine these extracted features with a convolutional neural network (CNN) to improve the classification accuracy of the CNN. The authors evaluate the performance of their approach using multiple datasets, such as MNIST and SVHN, and demonstrate the superiority in terms of high recognition accuracy, low memory requirements and fewer training iterations. The results show that the proposed approach outperforms the baseline CNNs, with an improvement of up to 10%. This approach could be used by research groups and computer vision systems that are concerned with accurate image recognition tasks.