Tea leaves are imperative to ensure tea qualities; like flavors and aromas. Although the challenge in differentiating healthy and diseased leaves results in major yield losses annually. To address this issue, researchers conducted numerous studies, yet lacked sufficient accuracy. This study assesses the efficacy of both supervised and semi-supervised learning approaches in tea leaf classification. Employing a custom dataset from Udalia Tea Garden, Bangladesh, DenseNet121 acquired the highest accuracy of 97%. By bridging the best supervised model with the FixMatch algorithm, the semi-supervised model achieved comparable performance of 96% utilizing only 25% lebeled data. These findings indicate that although supervised learning models offered superior accuracy, the performance of semi-supervised models holds a significant potential, specifically with small size of labeled data. To make quality tea, only the upper surface or young leaves are adopted and in this regard, our model serves a noble purpose by ensuring this specification. Moreover, this model stimulates a notable improvement in quality assurance mechanisms in high-quality tea production.