Abstract: As a common pharmaceutical container in the pharmaceutical industry,the quality of capsules is closely related to the therapeutic effect of the disease and the health status of the patient.The product quality inspection in the quality control process is of great importance to its mass production and practical utility.To realize the pixel-level segmentation of surface defects on pharmaceutical capsules,a surface defect detection method for pharmaceutical capsules based on feature fusion and semantic guidance was proposed.The feature fusion module was used to aggregate multi-scale semantic information to enable effective utilization of features at various levels for enhancing the segmentation ability of multi-class objects and tiny defects.To alleviate the issue of feature lose during coding and decoding,the semantic information was correctly channeled through the semantic guidance module to enhance the local effect of defect segmentation.Finally,the segmentation details of surface defects were further optimized under the role of the refine segmentation module.The evaluation results on the capsule defect dataset showed that the proposed method had a more balanced overall performance under multi-dimensional evaluation metrics including accuracy,speed,model size,and training time compared with existing methods.

Key words: deep learning, semantic segmentation, defect detection, attentional guidance, multi-scale fusion

摘要： 作为制药产业中常见的药剂容器,胶囊质量与病症的治疗效果以及患者身体的健康状况密切相关。因此,在胶囊生产质量管理流程中,产品质检技术对其批量生产与实际效用有着重要意义。为实现药用胶囊表面缺陷的像素级分割,提出一种基于特征融合与语义引导的药用胶囊表面缺陷检测方法。首先,利用特征融合模块聚集多尺度语义信息,使各级特征得到有效利用,以增强对多类目标以及细小缺陷的分割能力;其次,为缓解编解码过程中特征丢失问题,通过语义引导模块对语义信息进行正确疏导,提升缺陷分割的局部效果;最后,在细化分割模块的作用下,进一步优化表面缺陷的分割细节。在胶囊缺陷数据集上的评估结果表明,相比于众多现有方法,所提方法在多维度的评价指标下(包括精度、速度、模型大小以及训练时长)具有更为平衡的整体性能。

关键词: 深度学习, 语义分割, 缺陷检测, 注意力引导, 多尺度融合