基于改进FoveaBox的废杂塑料检测
2022,30(4):66-71
摘要：塑料制品的产量和种类的飞速增长，给废杂塑料的回收带来极大的挑战；目前仍然依靠大量人工分拣，面对恶劣和高强度的工作环境无疑亟待自动化升级；为解决上述问题，提出了一种改进的FoveaBox目标检测算法；针对废杂塑料分选背景复杂的问题，采用ResNeXt-101作为主干网络替代ResNet-50来提高特征提取能力；针对外形差异大的问题，采用带缩放系数的可变形卷积来提高卷积过程的有效感受野；针对目标间彼此遮挡问题，采用带层级控制因子的软化加权锚点机制来提高紧挨目标的检测精度；结果表明，基于改进FoveaBox的废杂塑料检测算法检测平均精度均值达到85.79%，检测速度为71.4ms，具有较强的实用性。
关键词：塑料分选;目标检测;改进FoveaBox;可变形卷积;软化加权锚点
Detection of Waste Plastic Based on improved FoveaBox
Abstract：The rapid growth in the output and types of plastic products has brought great challenges to the recycling of waste and miscellaneous plastics. At present, it still relies on a large number of manual sorting, facing the harsh and high-intensity working environment, it is undoubtedly urgent to upgrade automation. To solve the above problems, an improved FoveaBox target detection algorithm is proposed. In view of the complicated background of waste plastic sorting, ResNeXt-101 is used as the backbone network to replace ResNet-50 to improve the feature extraction ability. Aiming at the problem of large shape differences, a deformable convolution with a zoom factor is used to improve the effective receptive field of the convolution process. Aiming at the problem of mutual occlusion between targets, a softened weighted anchor point mechanism with hierarchical control factors is used to improve the detection accuracy of close targets. The results show that the mean average precision of the waste plastic detection algorithm based on the improved FoveaBox reaches 85.79%, and the detection speed is 71.4ms, which has strong practicability.
Key words：plastic classification; target detection; advanced FoveaBox; deformable convolution; soft-weighted anchor point
收稿日期：2021-10-15
基金项目：国家重点研发计划(2019YFC1908201)
