Image-Based Essay Scoring Deep Learning Using a CNN Model GoogLeNet

Automatic essay assessment is a crucial aspect of evaluating students' academic performance, but the manual grading process by educators is often time-consuming and subjective. This paper proposes using a deep learning application for automatic essay assessment by leveraging GoogLeNet Convolutional Neural Network(CNN) for handwriting analysis. The system aims to alleviate the workload of educators by allowing them to upload answer sheets for automatic grading. Students can access their results through a separate application. The system is designed to handle various handwriting styles through a preprocessing stage that involves cleaning answer sheets, extracting answer boxes, and removing empty spaces. Labelled data is then used to train the GoogLeNet model, which analyzes the answer sheets and provides scores based on student data containing 12 answer keys. The final scores are returned to students via the application. This study employs a 60 % and 40 % ratio for testing and training data, which yielded the best performance after experimenting with other ratios. Additionally, experiments were conducted to optimize hyperparameters, resulting in a learning rate of 0.002, Adam optimizer, and batch size of 32. Grading was performed for each question number, as they have different datasets, with an average accuracy of 85%. The lowest result was obtained for question 3b with a percentage of 45.83 %, while the highest result for question 3c was 100%. In conclusion, the use of GoogLeNet CNN for automatic essay assessment shows significant potential for improving the efficiency and accuracy of academic grading, although accuracy still varies across different question numbers.