Abstract: In recent years,there are strict requirements on polarizer’s production quality with its application more widely.The target detection algorithm of deep learning was adopted to detect three kinds of defects of polarized,which solved the problems of low detection accuracy and high hardware cost and optimize the production process.Based on YOLOV3-Tiny algorithm,Dense Block module and SPP-Net module were used to optimize the feature extraction network,and combines with the actual situation of the target to be detected to adjust the optimized network detection module,thus an improved algorithm YOLOV3-Tiny-D was proposed.The experiment showed that when the method was tested on polaroid data set,the detection accuracy of three kinds of defects namely dirt,scratch and mark were 90.74%,98.90% and 97.52% respectively,while the detection speed was guaranteed at the same time,and the average accuracy was 95.72%,7% higher than the original algorithm.

Key words: polarizer, defect detection, SPP-Net module, Dense Block module, YOLOv3-Tiny-D algorithm

摘要： 随着偏光片的应用日益广泛,对于其生产质量的要求也愈加严苛。采用深度学习的目标检测算法对偏光片的三类瑕疵缺陷进行检测,以解决传统方法检测精度低、硬件成本高的问题,从而优化生产工艺。基于YOLOv3-Tiny算法,采用Dense Block模块与SPP-Net模块对其特征提取网络进行优化,并与待检测目标的实际情况相结合调整优化网络的检测模块,提出一种改进后的算法YOLOv3-Tiny-D。实验表明,所提方法在偏光片数据集上测试时,单张图片在保证检测速度的同时(18ms/张),脏污、划痕、标记3类缺陷的检测正确率为9074%、9890%、9752%,平均正确率9572%,较原算法提高7%。

关键词: 偏光片, 缺陷检测, SPP-Net模块, Dense Block模块, YOLOv3-Tiny-D算法