Research on a Product Quality Monitoring Method Based on Multi Scale PP-YOLO

To monitor product quality in the production process real time, this thesis proposes a monitoring model based on PaddlePaddle You Only Look Once (PP-YOLO). First, preprocessing stage, data enhancement method and K-means++ are used to improve robustness of algorithm, generated anchor box can screen more refined features earlier. Second, ResNet50-vd with deformable convolution idea is selected as backbone detection model, feature pyramid network structure composition loss function improved, learning ability enhanced enable it detect multiple scales defects. Finally, pruning performed basis trained reduce number parameters so that be deployed industrial scenarios limited hardware conditions. Experimental results show proposed meet requirements for speed accuracy actual production, providing new concept deployment deep models field.