基于RB-YOLO的生活垃圾实时检测算法

Real-time detection algorithm of domestic garbage based on RB-YOLO

生活垃圾图片背景复杂,同一类垃圾特征差异性大,给垃圾检测带来了极大的挑战。为实现复杂场景下生活垃圾的实时检测,论文提出了一种基于改进YOLOv3的生活垃圾检测算法RB-YOLO。在该垃圾检测模型中,将YOLOv3与轻量级卷积神经网络RepVGG相结合,解耦模型训练和推理,在保证特征提取能力的同时提高模型推理速度。采用加权双向特征金字塔网络BiFPN优化原有的特征金字塔结构,通过双向特征融合充分利用多尺度语义特征。同时引入空洞空间金字塔池化ASPP增强特征提取,获取多尺度特征信息。在提出的DGD数据集上进行实验分析,与YOLOv3模型相比,RB-YOLO模型在垃圾检测任务中准确率和速度都有所提升,同时在测试集和网络图片上也取得了不错的检测效果,验证了RB-YOLO的有效性和鲁棒性。

The background of domestic garbage pictures is complicated,and the large differences in the characteristics of the same type of the garbage have brought great challenges to garbage detection.In order to realize the real-time detection of domestic waste in complex scenarios,the paper proposes a domestic waste detection algorithm RB-YOLO based on improved YOLOv3.In this garbage detection model,YOLOv3 is combined with the lightweight convolutional neural network RepVGG to decouple model training and reasoning,and improve the speed of model reasoning while ensuring feature extraction capabilities.The weighted bidirectional feature pyramid network BiFPN is used to optimize the original feature pyramid structure,and multi-scale semantic features are fully utilized through bidirectional feature fusion.At the same time,the hollow space pyramid pooling ASPP is introduced to enhance feature extraction and obtain multi-scale feature information.Based on the above research,experimental analysis is performed on the proposed DGD data set.Compared with the YOLOv3 model,the RB-YOLO model has improved accuracy and speed in garbage detection tasks.Furtherly,it has also achieved good detection results on the test set and network pictures,and verified the effectiveness and robustness of RB-YOLO.