用于智能垃圾分选的轻量级检测算法

为实现垃圾分选自动化,确保垃圾正确分类,提出了一种基于YOLOv4的轻量级垃圾检测算法.算法对YOLOv4中的主干网络CSPDarknet53,使用层级调整后的MobileNetV3网络进行替换,使得网络架构更适用于YOLOv4网络,并提升网络的检测速度;同时结合Ghost模块和MobileNeXt网络结构思想,设计了一种全新的bottleneck,用以替换主干网络中的bottleneck,以提升模型的检测精度;接着在主干网络中添加大残差边结构,以提升网络的检测精度;然后在颈部网络之前添加CA (coordinate attention)注意力机制,进一步提升网络的检测精度;最后为避免K-means算法在聚类过程中陷入局部极值,使用二分K-means算法对垃圾检测数据集进行anchor box的重新聚类.实验结果表明,重新设计的网络与YOLOv4网络的mAP值相近,但参数量减少了89%,检测速度提升了51%, FPS值达到了67.5 (on NVIDIA GeForce RTX 3060),可实现部署到算力和内存较低的嵌入式设备中.

Lightweight Detection Algorithm for Intelligent Garbage Sorting

To achieve automatic garbage sorting and ensure correct garbage classification, this study proposes a lightweight garbage detection algorithm based on you only look once version 4 (YOLOv4). The algorithm replaces CSPDarknet53, the backbone network in YOLOv4, with a MobileNetV3 network after level adjustment, thereby making the network architecture more suitable for the YOLOv4 network and improving the detection speed of the network. Subsequently, a new bottleneck is designed by drawing on the idea of the Ghost module and the MobileNeXt network structure to replace the bottleneck in the backbone network and boost the detection accuracy of the model. Then, a large residual edge structure is added to the backbone network to raise the detection accuracy of the network. Furthermore, a coordinate attention (CA) mechanism is added before the neck network to further enhance the detection accuracy of the network. Finally, to prevent the K-means algorithm from falling into a local extremum in the clustering process, this study employs the bisecting K-means algorithm to re-cluster anchor boxes for the garbage detection dataset. Experimental results show that the redesigned network achieves a mean average precision (mAP) close to that of the YOLOv4 network but reduces the number of parameters by 89%. Moreover, it improves detection speed by 51% to 67.5 FPS (on NVIDIA GeForce RTX 3060). It can thus be deployed to embedded devices with low computing power and memory.