面向边缘计算设备的改进型YOLOv3垃圾分类检测模型

Improved YOLOv3 Garbage Classification and Detection Model for Edge Computing Devices

为提高垃圾分类的自主化和智能化程度,垃圾桶需要配备视觉传感器和搭载有效的垃圾检测与分类算法的智能硬件。针对该需求,提出了一种基于改进型YOLOv3的智能化垃圾识别分类算法。首先,引入MobileNetv3网络代替YOLOv3的主干网络Darknet53,并加入空间金字塔池化结构,在减少网络模型计算复杂度的同时保证模型准确率;其次,采用4个不同的尺度检测加强模型对小目标的检测能力;然后,采用complete intersection over union(CIOU)损失函数替换原有YOLOv3模型的损失函数,进一步提升模型的精确度;最后,搭建家用垃圾桶测试平台,并将所提算法移植到边缘计算模块NVIDIA Xavier NX上。实验结果表明,所提轻量化优化算法在服务器和NVIDIA Xavier NX平台上对于自主建立的垃圾数据集平均精度一致,达到了72.1%,比YOLOv3提高了4.9个百分点,比YOLOv4略低1.6个百分点;检测速度分别达到了74,19 frame/s,远高于YOLOv3算法的43,8 frame/s及YOLOv4算法的50,11 frame/s,表明所提算法满足边缘计算设备的要求,具备潜在的应用价值。

In order to promote the degree of autonomy and intelligence of garbage classification, the garbage bin need be equipped with visual sensor and intelligent hardware carrying effective garbage detection and classification algorithm. To meet this demand, an improved garbage identification and classification algorithm based on YOLOv3 is proposed. First, MobileNetv3 network is introduced to replace Darknet53, the backbone network of YOLOv3,and spatial pyramid pooling structure is added to reduce the computational complexity of the network model and ensure the accuracy of the model. Second, four different scales are used to enhance the detection ability of the model to small targets. Then, the loss function of the original YOLOv3 model is replaced by the complete intersection over union(CIOU) loss function to improve the accuracy of the network model. Finally, a household trash can test platform is built, and the proposed algorithm is transplanted to the edge computing module NVIDIA Xavier NX.The experimental results show that the average accuracy of the proposed optimization algorithm is consistent on the server and NVIDIA Xavier NX platform in the self-made garbage dataset,reaches 72. 1%, which is 4. 9 percentage point higher than that of YOLOv3 and 1. 6 percentage point lower than that of YOLOv4;detection speed is74, 19 frame/s, which is much higher than 43, 8 frame/s of YOLOv3 algorithm and 50, 11 frame/s of YOLOv4 algorithm, indicating that proposed algorithm meets the requirements of edge computing equipment and has potential application value.