Image-based detection and classification of screws and nuts using deep learning

Resumo

Object detection in images has been one of the biggest challenges for computer vision researchers. This
paper presents a case study on screws and nuts detection and classification. The identification of screws is not a
trivial task. There are about 1,500 unique bolts, and in some cases, the differences between the two pieces involve
only tiny details, making identification difficult for untrained people. The experiments used an MVTec Screws
dataset with 384 images of bolts and nuts on a wooden background. The classification problem consists of 10
classes that differ in the length and width of the screws or nuts diameter, the color of the metal, and the shape
of the screw head, tip, or thread. Objects in some images are separated, in others ones, objects are together or
overlapped. For screws and nuts detection and classification, the network YOLOv4 and the Darknet framework
were used for training and inference. Performance was evaluated considering detection and classification after
4,200 epochs run. The results in detection, in terms of IoU and mAP, scored 77.79% and 97.79%, respectively. In
classification tasks, all classes reached above 99% F1-Score.