DOI：10.3969/j.issn.1672-3872.2024.03.012
基金项目：2022年国家级大学生创新创业训练计划项目“猕猴桃特征提取与自动分级关键技术的研究”（202210671147）
作 者：郝家英 ，曾玉娇 ，王祥文 ，陈 园 ，刘文江 （贵州财经大学信息学院，贵州 贵阳 550025）
摘 要：【目的】当前猕猴桃的缺陷识别与分级大多基于传统图像处理技术，无法完全脱离人工操作，且对图像的质量要求高、可靠性差，无法满足猕猴桃现有的分级需求。【方法】课题组提出了一种基于Deeplabv3+网络模型的猕猴桃特征提取和自动分级方案，采用轻量级卷积神经网络MobileNetV2作为图片特征提取工具，将采集的466张猕猴桃图片进行滤波处理、数据增强等步骤后获得2 796张图片，按8∶2的比例分为训练集和预测集进行网络模型训练，并与传统模型识别率进行了对比。【结果】基于Deeplabv3+网络模型的猕猴桃四个等级的识别率分别为100%、96.15%、95.83%和97.05%，高于传统计算机视觉方法的识别率，验证了该方法在猕猴桃分级上的可行性。【结论】该模型有效地降低了系统的参数和计算量，具有训练时间短、空间复杂度低等优点，在猕猴桃图像的特征抽取和分级任务上具有较好的表现。
关键词：猕猴桃；自动分级；特征提取；Deeplabv3+
Study on Feature Extraction and Automatic Classification of Kiwifruit Based on Deeplabv3+ Network
Hao Jiaying, Zeng Yujiao, Wang Xiangwen, Chen Yuan, Liu Wenjiang
(School of Information, Guizhou University of Finance and Economics, Guizhou Guiyang 550025)
Abstract: [Objective] Currently, the defect recognition and grading of kiwifruit are mostly based on traditional image processing techniques, which cannot be completely separated from manual operations. In addition, there is a high demand for image quality and poor reliability, which cannot meet the existing grading requirements of kiwifruit. [Method] The research group proposed a kiwifruit feature extraction and automatic classification scheme based on the Deeplabv3+network model. The lightweight convolutional neural network MobileNetV2 was used as the image feature extraction tool, and the collected 466 kiwifruit images were filtered, data enhanced, and other steps were taken to obtain 2 796 images. The network model was trained by dividing it into a training set and a prediction set in an 8:2 ratio, And compared with the recognition rate of traditional models. [Result] The recognition rates of kiwifruit four grades based on Deeplabv3+network model were 100%, 96.15%, 95.83% and 97.05%, respectively, which were higher than the recognition rates of traditional computer vision methods, verifying the feasibility of this method in kiwifruit grading. [Conclusion] This model effectively reduces the parameters and computational complexity of the system, and has advantages such as short training time and low spatial complexity. It performs well in feature extraction and classification tasks of kiwifruit images.
Keywords: kiwifruit; automatic classification; feature extraction; Deeplabv3+
引文信息：[1]郝家英，曾玉娇，王祥文，等.基于Deeplabv3+的猕猴桃特征提取和自动分级研究[J].南方农机，2024，55(3)：49-54.
查看全文请下载PDF文件↓
