Fin Coil Dent Detection Using Deep Learning

Abstract

This work introduced the object detection using deep learning in order to detect and locate dents on fin coils. It was aimed at using the AI as a tool to detect fin coil dents and restore them in the fin coil manufacturing and assembly processes. Not only were the images of dents used to train object detection models, two different types of distinctive marks were added for the purpose of positioning calibration in the system. Three scalable models of the state-of-art EfficientDat D0, D1 and D2 were used and compared for their accuracies and performances. All models were trained successfully with the custom dataset. It took only 30 epochs to achieve a functionable performance. The dent detection accuracies which were considered from the True Positive, obtained from the D0, D1, and D2 models were <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$55 \%, 66 \%$</tex> and 75 % respectively. The three models can identify additional marks with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 0 0 \%}$</tex> accuracy. In all models, there was no False Negative detection in all object classes showing good potential of using the models in the real applications.