Deep Learning-Based Marker Classification And Detection For Augmented Reality In Ethiopian High School Biology Education

Abstract

In Recent Decades, The Significance Of Augmented Reality (AR) In Its Application For Several Fields Of Study Has Been Taken Into Consideration. The Same Could Also Be Stated For Deep Learning (DL). The Combined Effort Of These Two Concepts Has Been Deemed Cruciallyimportant In Medicine, Education, The Military, Marketing, Games, Entertainment, Sports, Etc. Applying DL In Augmented Reality Marker Detection On Educational Books Is The Focus Area Of This Research. This Study Aims In Finding An Efficient Deep-Learning Solution Suited For Marker Classification And Detection Of Augmented Reality Based On Ethiopian Educational Books.Using A Pre-Trained Model Is The Best Option For Marker Classification Due To The Low Amount Of Data Available. The Pre-Trained Models Used As Feature Extractors For Classification Include Mobilenetv2 And Mobilenetv3. The Training On Detection Has Been Done With The Yolov7 Detection Algorithm. Secondary School Books Published By The Ministry Of Education (MOE)Are Chosen For This Study, Specifically Biology From Grades 9 To 12 Since It Is A Subject That Has Ample Abstracted Figures That Require More Explanation And Demonstration Than A Simple Description With Words. The Dataset Gathered From Each Book Is By Choosing At Most 20 Figuresas Markers And Building The Dataset From There, By Keeping Several Instances Of The Environment In Mind, And To Help Enrich The Dataset, A Synthetic Dataset Is Used. The Overall Dataset Size Consists Of Over 17,000 Images Created Using The Book Figures For The Training. The Proceeding Process Has Been Preprocessing And Feature Extraction Then Was Followed By The Deep Learning Process Which Is To Train The Enhanced Data. For Classification,Mobilenetv3large Has Performed Better With 96.21% Accuracy Than The Rest Of The Pre-Trained Models Used As Feature Extractors. The Best Model For Detection Has Been Used To Create An AR System That Overlays Information On Top Of The Detected Markers. In Comparison, We Have Achieved A Detection With Better Performance And Higher Inference Speed By Using The Proposed Architecture Than The Previous Works.