【摘要】 为了提高深度学习在X射线注塑工件缺陷检测中的准确性，实现更高精度的无损检测，提出了一种改进的YOLOv8-seg注塑件内部缺陷实例分割模型PIS-YOLO。首先为了减少参数量并提高特征融合能力，在主干网络中设计了一个多尺度特征融合与通道数减小的HG-Net模块，取代传统C2f模块；进一步引入倒置残差移动块融合高效多尺度注意力(iRMB＿EMA)模块增强了深层传递，经由简化冗余连接的路径聚合网络-特征金字塔网络(PAN-FPN)完成特征融合。同时增加一个额外的输出分割检测头捕捉细小缺陷，提高了模型对小目标缺陷及缺陷边缘的精确识别。在注塑工业零件自制数据集上，主干网络部分提出的HG-Net模块对比C2f模块实现了同架构下计算量减小22.03%,在此基础上，结合iRMB＿EMA注意力融合模块与额外输出检测头设计的模型整体的准确率相较于基准模型分别提升了2.9%和5.7%,且模型更轻量，计算复杂度更低。更多还原

【Abstract】 To improve the accuracy of deep learning in X-ray injection molding workpiece defect detection and realize higher precision nondestructive testing, an improved YOLOv8-seg internal defect segmentation model PIS-YOLO was proposed in this paper. Firstly, to reduce the number of parameters and improve the feature fusion capability, a multi-scale feature fusion and channel number reduction HG-Net module is designed in the backbone network to replace the traditional C2f module. The iRMB＿EMA attentional fusion module is further introduced to enhance the deep transmission, and the feature fusion is completed by PAN-FPN with simplified redundant connections. Meanwhile, an additional output segmentation detection head is added to capture small defects, which improves the model’s accurate recognition of small target defects and defect edges. On the self-made data set of injection molding industrial parts, HG-Net module proposed in the backbone network section achieves a 22.03% reduction in computation under the same architecture compared with C2f module. On this basis, the overall precision of the model combined with the iRMB＿EMA attention fusion module and additional output detection head is improved by 2.9% and 5.7%, respectively, compared with the benchmark model, and the model is lighter and less complex.更多还原