Surface defect detection on bolt surface using a real-time fine-tuned YOLOv6 model

Assessing metallic components for surface imperfections is essential for ensuring product quality, enhancing manufacturing efficiency, and reducing expenses. Detecting minor surface imperfections is essential for guaranteeing the reliability and safety of metal components in operation. The existing methods for detecting minute imperfections on metal surfaces exhibit limitations, such as inadequate real-time performance and diminished accuracy. This study introduces an innovative real-time model, YOLOBolt, for bolt surface identification using a finetuned version of You Only Look Once (YOLOv6). The proposed model is termed YOLOBolt as it helps in identifying the defects on the surface of bolts. The Spatial pyramid pooling (SPP) block of the backbone of the baseline YOLOv6 model is replaced with a residual block. This mitigates semantic loss, reduces information loss, and eliminates the low-resolution feature layer. The backbone design of the proposed YOLOBolt employs the Hybrid Extraction of Features Algorithm (HEFA) for feature extraction. Additionally, YOLOBolt employs the Convolutional Block-Attention Mechanism (CBAM) to enhance the model's precision in detecting small-sized flaws. Soft Intersection over Union (SIoU) is utilized to augment target detection. The model was trained using a real-time dataset from the bolt manufacturing sector in Panipat, India. The dataset comprises hexagonal bolts constructed from a prevalent metallic element. The suggested model demonstrated a 6.7% enhancement over the baseline Yolov6 regarding mAP (mean average precision), attaining a mAP of 96.50%. The results indicate that the proposed YOLOBolt enhances the efficacy of real-time identification of surface defects in metal components.