摘要
现阶段我国主要靠人工对垃圾进行分拣,存在安全系数低、效率低下等问题。传统目标检测方法针对种类繁多,形态各异的垃圾目标不易设计特征,鲁棒性较差,为实现自然环境下垃圾的快速精准识别,本文提出一种基于深度学习的轻量级垃圾分类检测方法。该方法通过引入CIOU边框回归损失函数来提高回归框准确率;针对低功耗移动设备终端的部署,提出一种以YOLOv3目标检测算法为基础,结合MobileNetV3的特征提取网络,对算法进行轻量化;在YOLO层加入GRU结构,利用多门控循环神经网络结构对YOLO层中不同大小的特征图建立记忆链接,对深层语义特征的向前融合过程进行过滤和筛选,使得特征融合效果更佳;使用迁移学习预训练的方式来提高模型的特征提取能力和泛化能力。文本采用自制的Garbage数据集对改进后的网络进行训练和测试,结果表明,本文提出的算法识别效果显著,平均准确率为90.50%,高于原YOLOv3网络的平均准确率86.30%,检测速度达到18帧/秒,满足实时检测的需求。实验表明,改进后的网络模型能在保证检测准确率和速度的同时,有效降低模型参数量,具有一定应用价值。
At present,China mainly relies on manual sorting of garbage,which has problems such as low safety factor and low efficiency.Traditional target detection methods are difficult to design features for a wide variety of garbage targets with different shapes,and have poor robustness.In order to achieve rapid and accurate garbage identification in natural environments,this paper proposes a lightweight garbage classification and detection method based on deep learning.This method improves the accuracy of the regression frame by introducing the CIOU frame regression loss function;for the deployment of low-power mobile device terminals,a YOLOv3 target detection algorithm is proposed based on the feature extraction network of MobileNetV3 to lighten the algorithm;The GRU structure is added to the YOLO layer,and the multi-gated recurrent neural network structure is used to establish memory links for feature maps of different sizes in the YOLO layer,and to filter and screen the forward fusion process of deep semantic features to make the feature fusion effect better;use Transfer learning pre-training is used to improve the feature extraction ability and generalization ability of the model.The text uses the self-made Garbage data set to train and test the improved network.The results show that the algorithm proposed in this paper has a significant recognition effect,with an average accuracy rate of 90.50%,which is higher than the original YOLOv3 network′s average accuracy rate of 86.30%,and the detection speed is up to 18 frames per second,meeting the needs of real-time detection.Experiments show that the improved network model can effectively reduce the amount of model parameters while ensuring the detection accuracy and speed,which has certain application value.
作者
许伟
熊卫华
姚杰
沈云青
XU Wei;XIONG Wei-hua;YAO Jie;SHENG Yun-qing(Zhejiang Sci-Tech University,Faculty of Mechanical Engineering and Automation,Zhejiang 310018,China)
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2020年第9期928-938,共11页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61803339)
浙江省重点研发计划项目(2019C03096)资助项目。
