Automatic Handwritten Answer Script Evaluation System using Improved Optimization in Hybrid Deep Residual Neural Network

Evaluating students answer scripts in today's educational systems is difficult since there are multiple evaluation standards, various types of questions and diverse approaches to the same issue. It can be challenging to guarantee impartiality in assessments because traditional manual evaluation is frequently uneven, ineffective, and occasionally biased. Dealing with various answer types makes these difficulties much more difficult. Because each one of these needs' different strategies, manual evaluation becomes more labor-intensive and prone to mistakes. This study suggests an Automatic Handwritten Answer Script Evaluation System (AHASES) to address these issues and reduce the workload for teachers by grading a sizable corpus of student answers. This solution leverages and integrates combination of Convolutional Neural Networks (CNN), Deep Residual Neural Network (ResNet101) and Connectionist Temporal Classification (CTC), while exploiting Word Beam Search (WBS) for decoding. A popular learning approach for training DNN models in a variety of disciplines is the Adaptive Moment Estimation (Adam) algorithm, a highly effective adaptive optimization algorithm. Its generalization performance needs to be improved, nevertheless, particularly when using sizable corpus of student answers dataset for training. As a result, the regular Adam algorithm and the AMSGrad method used to generate the modified hybrid mechanism known as the Improved Adam Optimizer (IAO) integrated in the CNN-ResNet101-CTC-WBS model has been suggested for AHASES in this research. The proposed IAO implemented in integration of a CNN-ResNet101-CTC-WBS model has shown great promise, achieving an impressive accuracy rate of 97.5% for performing AHASES.