基于改进的RetinaNet医药空瓶表面气泡检测

Bubble detection on the surface of medical empty bottles based on improved Retina Net

摘要:

医药空瓶在生产过程中瓶身表面会产生大量的气泡缺陷，但现有的方法对医药空瓶表面气泡检测存在各种问题，例如对复杂场景变化的鲁棒性不强，抗噪声干扰能力弱等.针对现有医药空瓶表面的气泡缺陷，提出了一种改进的深度学习目标检测算法RetinaNet对瓶身气泡进行检测.对原始RetinaNet算法中的特征金字塔网络结构进行了优化，在特征融合过程中引入了特征增强模块，用来提高网路对图像语义特征的提取，增强网络特征提取能力.为了减少模型的参数数目和计算时间，考虑到空瓶表面气泡均为小目标缺陷，去掉原始特征金字塔网络中用于检测大目标的网络结构，提高了算法检测速度.通过对标准的ResNet50网络进行重新组合，并引进了膨胀卷积模块，扩大特征图感受野，提高了模型检测的精度.通过在注塑空瓶数据集上对本文的方法进行了验证，其准确率为99.72%，漏检率为0.12%，误检率为016%，mAP为99.49%，相比原始的RetinaNet的mAP提高了接近2.4%.

Abstract:

In the production process of medical bottles, a large number of bubble will be generated on the surface of the empty bottle body, but the existing methods have various problems in the detection of bubbles on the surface of empty medical bottles such as the week robustness to complex scene changes, and the poor antinoise ability and so on. Aiming to the existed bubble defects on the surface of medical empty bottles, an improved deep learning target detection algorithm RetinaNet is proposed to detect the bubbles on the bottle body. This paper mainly improves the feature pyramid network structure in the original RetinaNet algorithm, and introduces the feature enhance module in the process of the feature fusion, which effectively improves the extraction of semantic features and expands the receptive field of feature maps. In order to reduce the number of parameters and calculation time of the model, considering that the bubbles on the surface of the empty bottle are all small objects, the network structure to detect large objects in the original feature pyramid network is removed, which effectively improves the detection speed. By recombining the standard ResNet50 network, a dilation convolution module is introduced to expand the feature map receptive field and the accuracy of model detection is improved. The proposed method is validated on the empty dataset of injection molding, and the accuracy is 99.72%, the missed rate was 0.12%, the false detection rate was 0.16%, the mAP is 99.49% which is higher by nearly 2.4% compared with the original RetinaNet.