MSP-YOLO:用于识别水果的改进算法
MSP-YOLO:AnImproved Algorithm for Identifying Fruits
余鹏泽 1刘兴德 2谢延楠 1任洛莹 1孔志成 1胡文松1
作者信息
- 1. 吉林化工学院信息与控制工程学院,吉林吉林 132022
- 2. 吉林化工学院机电工程学院,吉林吉林132022
- 折叠
摘要
目标检测在计算机视觉中占有重要的地位,并逐渐成为许多应用的技术基础.将目标检测运用在水果识别中可以提高水果采摘的效率,由于对水果识别的背景复杂性、模型相似性大,纹理干扰严重、水果局部遮挡等问题,且基于传统方法的水果目标识别效率比较低.为了解决这一问题,提出了一种基于YOLOv5的水果检测分类算法MSP-YOLO.首先,将YOLOv5的主干网络换成另外一种更加轻量化的主干网络MobileNetV3,在减小模型大小的同时也能够提高对水果的检测速率;其次,将引入SE注意力机制到改进网络中,在YOLOv5基础模型的颈部中间层网络中加入SE注意力机制.SE注意力机制的优点在于可以帮助模型对通道信息最多的特征通道给予更多的注意力,来达到抑制那些对整体不太重要的通道特征,从而提升准确率;最后,通过将损失函数CIoU更改为MPDIoU简化了边界框之间的相似性比较,能更好地优化数据集,提高水果检测的识别精度.实验结果表明,MSP-YOLO在数据集上的精度达到92.7％mAP,比未改进的YOLOv5提高了 3.3％,改进的算法在检测精度和速率方面优于目前主要的几种目标检测模型 Faster R-CNN、SSD、YOLOv7-tiny、YOLOv3-tiny、YOLOv4,以及原模型 YOLOv5.
Abstract
Object detection occupies an important place in computer vision and is gradually becoming the technical basis for many applications.The application of object detection in fruit recognition can improve the efficiency of fruit picking,but the efficiency of fruit object recognition based on traditional methods is relatively low due to the problems of background complexity,large model similarity,serious texture interference,and local occlusion of fruits.In order to solve this problem,this paper proposed a fruit detection and classification algorithm MSP-YOLO based on YOLOv5.Firstly,the backbone network of YOLOv5 was replaced with another more lightweight backbone network,MobileNetV3,which could reduce the size of the model and improve the detection rate of fruits.Secondly,this paper introduces the SE attention mechanism into the improved network,and adds the SE attention mechanism to the neck middle layer network of the YOLOv5 base model.The advantage of the SE attention mechanism is that it can help the model pay more attention to the feature channels with the most channel information,so as to suppress those channel features that are not important to the whole,so as to improve the accuracy.Finally,by changing the loss function CIoU to MPDIoU,the similarity comparison between the bounding boxes is simplified,which can better optimize the dataset and improve the recognition accuracy of fruit detection.Experimental results show that the accuracy of MSP-YOLO on the dataset reaches 92.7％mAP,which is 3.3％higher than that of the unimproved YOLOv5,and the improved algorithm is superior to the main object detection models Faster R-CNN,SSD,YOLOv7-tiny,YOLOv3-tiny,YOLOv4 and the original model YOLOv5 in terms of detection accuracy and rate.
关键词
注意力机制/YOLOv5/水果检测/MobileNetV3/损失函数Key words
attention mechanism/YOLOv5/fruit detection/MobileNetV3/loss function引用本文复制引用
出版年
2024
NETL
NSTL
万方数据