Handwritten computer science words vocabulary recognition using concatenated convolutional neural networks

• Published in: Multimedia tools and applications Vol. 82; no. 15; pp. 23091 - 23117
• Main Authors: Hamida, Soufiane, El Gannour, Oussama, Cherradi, Bouchaib, Ouajji, Hassan, Raihani, Abdelhadi
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2023; Springer Nature B.V
• Subjects: Artificial neural networks; Colleges & universities; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Feature extraction; Handwriting; Handwriting recognition; Image classification; Multimedia Information Systems; Neural networks; Performance evaluation; Preprocessing; Special Purpose and Application-Based Systems; Words (language); Transfer-learning; Handwriting recognition; CNN; Features extraction; Concatenation technique
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-14105-2

Handwriting recognition is a multi-step process that includes data collection, preprocessing, feature extraction, and classification in order to create a final prediction. This process becomes more and more delicate when dealing with the scriptures of college or secondary school learners. The primary purpose of this research is to offer an improved model for classifying images of computer science words vocabulary written by learners. Indeed, the aim is to develop a reliable handwriting recognition system for the benefit of the educational field. The proposed recognition model based on the combination of four pre-trained CNNs models, namely ResNet50 V2, MobileNet V2, ResNet101 V2, and Xception. Our earlier established Computer Science Vocabulary Dataset (CSVD) is used to build and validate the proposed concatenated model. Then, we have applied preprocessing operations to reduce irregularities, like fuzzy letters and distorted undefined symbols. The proposed CNN model is trained on the concatenated features generated by the four pre-trained CNNs using a parallel deep feature extraction approach. To evaluate the performance of our recognition system, we have used different common evaluation measures. The average accuracy of the proposed system for handwritten words vocabulary is 99.97%, and the overall loss rate is 3.56%. In addition, these performances have been compared with alternative state-of-the-art models and better performance has been observed.