Dense material detection method based on deep learning
Industrial intensive material inspection has high requirements for counting accuracy.Traditional methods perform well in cases where the density of detected samples is relatively uniform and they do not overlap.However,in scenarios with uneven density distribution or overlapping samples,the issue of misrecognition becomes more severe.To enhance detection accuracy and counting precision,this study improves upon the HRNet architecture and proposes a self-attention multi-scale fusion model.The main model employs HRNet for feature interaction fusion among different resolution feature maps and enhances global feature extraction by adding a self-attention mechanism to high-resolution feature maps.Furthermore,to address situations where the inspection performance is poor for materials with few but large components,a dual-channel material size determination mechanism is introduced,utilizing the YOLO framework for material size classification detection.Lastly,the datasets used in this study are collected and annotated using X-ray non-destructive testing equipment.The proposed model achieves a prediction accuracy of 96.7%on this dataset,demonstrating improvement compared to other models.
NETL
NSTL
万方数据
