基于改进型YOLO v3的蔬菜识别算法-《郑州大学学报(工学版)》

文章信息/Info

Author(s):: Wei Hongbin; Zhang Duanjin; Du Guangming; Xiao Wenfu; School of Information Engineering, Zhengzhou University

Keywords:: Vegetable identification; K-means; convolutional neural network; feature pyramid; YOLOv3

摘要:: 针对超市的散装蔬菜区排队称重问题（称重设备能够自动识别蔬菜种类将有效地提高超市的运行效率），提出一种基于改进型 YOLOv3的蔬菜识别方法。首先，利用高清摄像头以及网络爬虫技术采集蔬菜图片；其次，通过 K-means 聚类分析得到 15组适应于蔬菜数据集的先验框；再次，采用一种新的边界框回归损失函数 DIoU 来提高检测任务的精度；最后，因蔬菜数据集中的大目标较多，通过增强特征提取网络，获取 5组不同尺度的特征构成特征金字塔从而实现蔬菜识别任务。改进型YOLOv3算法在测试集上的平均精度 mPA 达到 93.2％，识别速度是 35f·s^-1。该方法在保证实时检测目标的同时提升了识别的平均精度。

Abstract:: The queuing and weighing problem was common in bulk vegetable area of supermarket. If weighingequipment could automatically recognize vegetable, it would effectively improve the operational efficiency ofsupermarket. Therefore, a vegetable recognition method based on improved YOLOv3 was proposed. Firstly ,vegetable pictures were collected by using high-definition camera and web crawler technology. Secondly, 15groups of anchors suitable for vegetable datasets were obtained by K -means clustering analysis. Thirdly, a newbounding box regression loss function DIoU was proposed to improve the precision of detection task. Finally, asthere were many large objects in vegetable datasets, 5 groups of feature pyramids with different scales were ob-tained by enhancing feature extraction network to realize vegetable detection task. The mAP of the improvedYOLOv3 algorithm on the test dataset was 93. 2%, and the recognition rate was 35 fps. This method improvedthe recognition of mAP while guaranteeing real-time object detection.