In the field of machine learning, various models are used to train and test machines to perform a variety of tasks, including data analysis, computer vision, speech recognition, pattern matching, pattern recognition, and natural language processing. Computer vision is an important area where deep learning architectures, GANs, and transfer learning are applied. In the field of pattern matching and pattern recognition, research based on deep learning models has attracted significant attention from researchers due to their excellent performance. Pashto is an ancient and historical language spoken in Pakistan and Afghanistan. Optical character recognition (OCR) systems using Generative Adversarial Network (GAN) have been developed for several cursive languages, such as Chinese, Urdu, and Japanese, but no study has been done yet to recognize Pashto handwritten characters using the GAN network architecture. In this study, we used a Generative Adversarial Network (GAN) to generate additional images of handwritten characters in order to help improve the accuracy of a recognition system for handwritten characters. We used a dataset of 43,000 images representing 43 different classes of Pashto characters, which was collected from GitHub repositories and is available for other researchers to use. We employed two deep transfer models, ResNet50 and VGGNet16, to recognize the Pashto handwritten characters and evaluated their performance. The results showed that VGGent16 achieved 98.39% testing accuracy while ResNet50 achieved 99.24% testing accuracy. From the experimental result found that ResNet50 achieved a higher accuracy rate than VGGNet16 when tested on the dataset of Pashto characters. Furthermore, the accuracy rate achieved by ResNet50 was higher than the accuracy rates reported in previous studies on Pashto character recognition.