Comparative Analysis of Machine Learning Models in Hand-Drawn Circuit Components Detection

This study closely examines four major machine learning models with diverse architectures for their application in detecting electronic components within hand-drawn circuits. Pre-existing checkpoints of models were used for further domain specific training on custom dataset of hand drawn circuit images using Transfer Learning. The models under investigation include MobileNet SSD, YOLOv5xu, YOLOv8, and DETR transformer. For training, custom image dataset of 480 images with 1200 unique labels across 5 classes was generated. Inferencing and testing was performed on all models. Testing the power and efficiency of each model involves examining their approach to interpreting intricate hand-drawn diagrams. Various circuit designs, ranging from simple to complex, have been employed in these evaluations. The objective is to ascertain the strengths and limitations of each model through a comprehensive analysis of technical metrics. Such insights can aid practitioners in the field of Object Detection in selecting the most suitable methodology for their endeavors. Additionally, this paper proposes several enhancements to current techniques for identifying components within hand-drawn circuits