基于RetinaNet的可回收垃圾有向目标检测
RETINANET-BASED DIRECTED TARGET DETECTION FOR RECYCLABLE WASTE
张铮 1邱达河 1金子博 1薛波 1胡新宇1
作者信息
- 1. 湖北工业大学机械工程学院,武汉 430068
- 折叠
摘要
可回收垃圾分拣是垃圾处理厂的重要工作,目前人工垃圾分拣效率低,工作环境恶劣,分拣成本高,为实现垃圾分拣的 自动化,基于视觉的可回收垃圾自动检测研究具有重要意义.针对传统的水平框目标检测算法在检测时易丢失目标的方向信息,定位框重合现象严重,无法获取目标真实长宽,不利于后续分拣的缺点,提出基于RetinaNet的有向 目标检测算法,该算法基于RetinaNet网络进行改进,在检测头中添加角度预测模块,使用PSC角度编码器改善角度回归边界问题,引入Balanced L1 loss损失函数平衡简单样本和困难样本的梯度贡献,替换骨干网络为Swin Transformer以增强网络特征提取能力.带角度预测的网络,能更准确地定位垃圾,改进后的网络精度(mAP)达到78.4％,比原算法提高了12百分点,同时与其他角度编码器相比PSC的检测效果均优于其他方法.
Abstract
To realize the automation of waste sorting,the research of vision-based automatic detection of recyclable waste is of great importance.To realize the automation of waste sorting,the traditional horizontal frame target detection algorithm loses the directional information of the target during the detection,and the overlap of the positioning frame is serious so that the true length and width of the target cannot be obtained,which is unfavorable to the subsequent sorting.The algorithm is based on the improvement of the RetinaNet network,adding the angle prediction module in the detection head,using the PSC angle encoder to improve the angle return boundary problem,introducing the Balanced L1 loss function to balance the gradient contribution of simple and difficult samples,and replacing the backbone network with the Swin Transformer to enhance the feature extraction capability of the network.The network with angle prediction can locate the garbage more accurately,and the improved network accuracy(mAP)reaches 78.4％,which is 12 percentage points higher than the original algorithm,while the detection effect of PSC is better than other methods compared with other angle encoders.
关键词
有向目标检测/深度学习/垃圾识别/角度编码器Key words
directed target detection/deep learning/waste detection/angle encoder引用本文复制引用
出版年
2024
NETL
NSTL
万方数据