Introducing Urdu Digits Dataset with Demonstration of an Efficient and Robust Noisy Decoder-Based Pseudo Example Generator

In the present work, we propose a novel method utilizing only decoder for generation of pseudo-examples, which has shown great success in image classification tasks. The proposed is particularly constructive when data are limited quantity used semi-supervised learning (SSL) or few-shot (FSL). While most previous works have an autoencoder to improve performance SSL, using single may generate confusing pseudo-examples that could degrade classifier’s performance. On other hand, various models utilize encoder–decoder architecture sample can significantly increase computational overhead. To address issues mentioned above, efficient means generating by generator (decoder) network separately each class be effective both SSL and FSL. our approach, trained random noise, multiple samples generated decoder. Our generator-based approach outperforms state-of-the-art FSL approaches. addition, released Urdu digits dataset consisting 10,000 images, including 8000 training 2000 test images collected through three different methods purposes diversity. Furthermore, explored effectiveness on FSL, demonstrated improvement 3.04% 1.50% terms average accuracy, respectively, illustrating superiority compared current models.