基于Faster R-CNN的玻璃瓶缺陷检测技术

The glass bottle defect detection based on faster R-CNN

自动化检测玻璃瓶缺陷技术的实现,能够减少人力物力的需求量,提高玻璃瓶缺陷检测结果的准确性以及可靠性。将深度学习网络应用到玻璃瓶缺陷检测技术上,分别使用VGG16和Resnet101作为缺陷检测模型中Faster R-CNN的特征提取网络,在不同的锚框的尺度下对缺陷检测效果进行分析。实验结果表明,Faster R-CNN在本实验的玻璃瓶缺陷数据集上检测准确率尚可,但对于小目标缺陷检测效果不理想;改进后的Faster R-CNN在玻璃瓶缺陷检测上mAP最高提升了5.19%,同时在检测小目标缺陷时的鲁棒性更强。

The realization of automatic glass bottle defect detection can reduce the demand of manpower and material resources while also improves the accuracy and reliability of glass bottle defect detection results.In this paper,deep learning is applied to the glass bottle defect detection technology,in which VGG16 and Resnet101 are used as the feature extraction network of Faster R-CNN in the defect detection mode,respectively.The effect of defect detection is analyzed at different anchor scales.The experimental results show that in the glass bottle defect data set,the detection accuracy of Faster R-CNN is good,whereas the detection effect for small target defects is not satisfactory.The improved Faster R-CNN can raise mAP by up to 5.19%in the detection of glass bottle defects,which also shows better robustness in detecting the defects of small targets.