Efficientdet Application for Detection of Incorrect Assemblies in the Antenna Manufacturing Process

Abstract

This work aims to investigate and effectively apply the new deep learning methods for object detection from the family of EfficientDet-Lite algorithms for the detection of incorrect assemblies in the antenna manufacturing process. In the proposed methodology, stateof-the-art pre-trained models are exploited and fine-tuned in a real industrial dataset with limited samples. The defects considered in the dataset include metal cracks, plastic cracks, and housing imperfection. The proposed approach with EfficientDet-Lite2 can achieve an Average Precision (AP50) of 73.68%. A Yolov5-based model was also developed and extracted comparable results. Moreover, the Average Precision for each class was computed. For conducting the experiments, the TensorFlow Lite object detection API was employed. The results of this research work are promising and support further investigation of these types of deep learning for defect detection in manufacturing.