基于局部特征和视觉词袋模型的大豆叶片病害识别
Identification of soybean leaf disease based on local features and visual bag of words model
郑金松 1谷海红 1蒋庆刚 1赵靖杰 1王贤 1韩增光1
作者信息
- 1. 河南理工大学鹤壁工程技术学院,河南鹤壁,458030
- 折叠
摘要
病害检测对提高大豆作物产量至关重要.针对传统视觉法诊断大豆作物病害而导致病害识别效率和分类准确率不高的问题,提出一种基于局部描述符和视觉词袋技术以数据表征大豆叶片图像的分类算法,同时保留有关潜在疾病的视觉信息.采用SIFT、DSIFT、PHOW和SURF 4种算法对大豆叶片的霜霉病、锈病TAN和锈病RB进行分类识别.结果表明,局部描述符PHOW表现出最佳的分类识别结果,其正确分类率为96.25%.进一步研究PHOW在不同颜色空间下的大豆病害识别效果.结果表明,与灰度图像相比,使用HSV、Opponent颜色空间可有效提升对大豆叶片病害检测的正确分类率,其正确分类率分别可达99.83%和99.58%,验证采用局部描述符和视觉词袋技术识别大豆叶片病害方法的可行性和高效性,并为其他作物的病害识别提供一种通用的分类识别方法.
Abstract
Disease detection is crucial for improving soybean crop yields.In response to the low efficiency and accuracy of disease recognition and classification caused by traditional visual diagnosis methods for soybean crop diseases,a classification algorithm based on local descriptors and visual bag-of-words techniques was proposed to represent soybean leaf images,while preserving visual information about potential diseases.Four algorithms such as SIFT,DSIFT,PHOW,and SURF,were employed to classify and recognize soybean leaf diseases such as downy mildew,rust TAN,and rust RB.The results demonstrated that the local descriptor PHOW yielded the best classification and recognition results,with an accuracy rate of 96.25%.Further research on the recognition effects of PHOW in different color spaces revealed that,compared to grayscale images,the use of HSV and Opponent color spaces could effectively improve the correct classification rate of soybean leaf disease detection,reaching accuracy rates of 99.83%and 99.58%respectively.This validates the feasibility and efficiency of identifying soybean leaf diseases by using local descriptors and visual bag-of-words techniques,and provides a general classification and recognition method for the disease recognition of other crops.
关键词
大豆叶片/病害识别/局部描述符/视觉词袋/颜色空间Key words
soybean leaf/disease identification/local descriptors/visual bag of words/color space引用本文复制引用
基金项目
河南省科技攻关项目(212102310488)
河南省科技攻关项目(212102310550)
河南理工大学鹤壁工程技术学院自然科学一般资助项目(2023—ZRYB—001)
出版年
2024
国家科技期刊平台
NSTL
万方数据