<p>Modern industries are heavily reliant on steel, which acts as an indicator of the progress of a nation. Steel is essential to the economy due to its variety of applications. However, surface defects in the steel hot-rolling process can significantly compromise the quality of the final product. Traditional manual examination-based defect classification techniques are manual, arbitrary, and less precise. Modern computer vision techniques have emerged as a potential solution, but deep learning models require extensive labelled data, which is costly and labour-intensive. Semi-supervised techniques, such as pseudo-labelling, may not provide an optimal guarantee of accuracy, and active learning methods often require significant manual annotation efforts, which are resource-intensive. To address the challenges of defect classification in steel production, we present a novel deep learning framework that enhances accuracy and reduces manual labelling effort. Our approach integrates pseudo-labelling with the application of Monte Carlo dropout-based active learning. By employing this method, we identify the most uncertain samples for manual annotation and use pseudo-labelling for certain cases, incorporating human verification for greater reliability. This iterative process efficiently utilises unlabelled data, optimising model performance with less dependency on extensive labelled datasets. We have employed a simulated verification process using a controlled library, allowing proof of concept evaluation of the hybrid framework. Experimental results on the steel defect dataset demonstrate that the entropy based heuristics achieved up to 99.73% classification accuracy, effectively reducing human labelling effort while maintaining robust performance.</p>