Missing wire thread inserts in mechanical components pose critical safety risks, as undetected defects may lead to mechanical failures and downtime. High-resolution (HR) imaging captures defects but faces challenges: complex metallic backgrounds, low-contrast substrates, and small defect sizes impede reliable detection. Due to limited contextual reasoning, traditional convolutional networks often misclassify textural noise or overlook genuine anomalies. This paper proposes Defect Detection Yolov5 (Ded-YOLOv5), an efficient defect detection framework, to address these issues. The model integrates a Defect Detection Attenion (Ded-Att) module into YOLOv5, adaptively enhancing defect features by modeling channel-spatial dependencies and suppressing background interference. A hierarchical fusion mechanism correlates multi-scale cues, improving localization accuracy for minute defects. Experiments demonstrate that Ded-YOLOv5 reduces false positives from metallic textures and minimizes missed low-visibility defects. The method exhibits superior anti-interference capability against artifacts like uneven illumination and reflective glares compared to conventional approaches. These advancements establish Ded-YOLOv5 as a practical solution for mechanical manufacturing quality assurance.