Computer Vision-Based Structured Data Collection and Intelligent Grading of Paper-and-Pencil Homework

Abstract

Aiming at the problems of low efficiency in traditional paper-and-pencil homework grading and the difficulty of digitizing process data, this study proposes a solution to automate the collection and grading of free-format handwritten homework. To this end, a "cloud-edge-end" collaborative system architecture based on computer vision is proposed. The system first uses cascaded image preprocessing and deep learning semantic segmentation models to accurately analyze the homework layout and locate the question areas. It then employs handwriting recognition models trained with domain adaptation and formula recognition models with context perception of the question stem to complete the structural extraction of the answer content. Finally, combining rule matching and semantic similarity calculation, it achieves intelligent grading of both objective and subjective questions. Experimental results show that on the self-built real-world dataset, the proposed method significantly outperforms other methods in key tasks such as question area segmentation mIoU of 0.94, handwriting formula recognition accuracy of 86.4%, and objective question grading F1 score of 97.5%. It also demonstrates stronger robustness in dealing with challenges of image quality, layout complexity, and writing standardization, with an average performance degradation rate of only 11.3%. This study confirms that the proposed deep visual understanding approach can effectively tackle the key challenges of automated handwritten homework processing and provides an efficient and reliable tool for educational informatization in terms of data collection and intelligent grading.