Research on the Appearance Inspection Technology of Screw Assembly Quality Based on Machine Vision

Aiming at the demand for screw visual inspection in the appearance inspection of a certain complex product assembly quality, a real-time multi-scale small object detection and classification model (RMSDC) is studied. RMSDC uses YOLOv8n as the base network structure. SPD-Conv consisting of a space-to-depth (SPD) layer and a non-strided convolution layer is used to improve the performance of the model in low-pixel images and small target detection. A convolution attention mixing (CAMixing) block is introduced in the middle of the backbone network and the neck network, which combines the attention mechanism with the multi-scale convolution to enhance the multi-scale feature extraction capability and improve the model accuracy. Choose an ultra-lightweight and effective dynamic upsampler (DySample), which is lightweight and can preserve the high-level feature semantic information of small targets as much as possible. The deformable convolution v4 (DCNv4) is fused to the dynamic head (Dyhead) to improve the generalization ability of the model in terms of shape transformation. Through comparative experiments, RMSDC in this paper improves the recall (R), mAP0.5 and mAP0.5:0.95 by 3.7%, 1.5%, and 1.8% compared to the baseline model on the experimental dataset, and mAP0.5 reaches 98.5%. The excellent performance of RMSDC has also been demonstrated on the ROSD dataset.