Automated Data Extraction and Character Recognition for Handwritten Test Scripts Using Image Processing and Convolutional Neural Networks

Evaluating students through examination scripts in educational environments is crucial, particularly with 'Mastery Feedback' from educators, enhancing student understanding and self-regulation. However, it has remained a hectic exercise requiring some innovative solutions. This study proposes integrating robotics to automate recording and collating marked scripts to reduce the burden on lecturers and improve productivity. Key objectives include developing a data extraction pipeline using methods like Oriented FAST and Rotated BRIEF (O.R.B.) for image alignment and adaptive thresholding for lighting variations. Additionally, a character recognition model using a Single Input Convolutional Neural Network (SICNN) was designed with three preprocessing techniques—binarisation, thinning, and gradient magnitude calculation— tailored to different image requirements. Training on the 'EMNIST by_merge' dataset showed varied validation accuracies, with the gradient input SICNN model achieving the highest at 89.24% overall and the binary input SICNN model excelling with 99.39% on custom scripts. This approach aims to enhance educational administrative processes and efficiency and thus achieve sustainable education.