Recognition of Ophiopogon japonicus Disease Based on Image Feature Fusion

Tao YANG Jingjing MA

Abstract The images of three diseases of black spot， anthracnose and leaf blight in Sichuan Ophiopogon japonicus leaves were taken the research object， and the bimodal method， Otsu threshold segmentation method and K-means clustering segmentation algorithm were compared and analyzed on the images of O. japonicus. The segmentation effect showed that the K-means clustering algorithm combined with the mathematical morphology processing method could meet the segmentation requirements. Then， the color， shape， and texture information of the lesion image were extracted and fused into a feature vector. Next， the variance analysis and principal component analysis method were applied to eliminate the feature parameters with poor disease characterization ability and reduce the eigenvector dimension to 10 dimensions. Finally， the classifiers for disease identification were designed by the support vector machine， and the recognition rate reached 90% after testing. The method has the advantages of low cost， simple algorithm and high efficiency， and basically meets the requirements of practical applications.

Key words Ophiopogon japonicus; Image processing; PCA; SVM; Disease recognition

Received： March 1 2021  Accepted： May 2 2021

Supported by Chengdu Agricultural College Funded Project （20ZR108）.

Tao YANG （1991-）， male， P. R. China， teaching assistant， master， devoted to research about machine vision and its applications.

*Corresponding author. E-mail： egstao@163.com.

The "Hometown of Ophiopogon japonicus" in China， Santai County， has more than 500 years of O. japonicus planting and cultivation history. "Fucheng O. japonicus" has become a well-known brand of O. japonicus in China， and its output accounts for more than 60% of the country and its export volume accounts for more than 80% of the country. And the output value exceeds 1.5 billion yuan[1]. However， the whole production process of O. japonicus is not highly mechanized， and mainly depends on manpower. Especially in the prevention and control of O. japonicus disease， over-reliance on farmers planting experience， lack of theoretical basis and scientific guidance， often missing the best disease control period or improper control measures brought huge economic losses[2-3]. The prevention and control of crop diseases has always been one of the hot topics studied by scholars. Sun et al.[4] used the convolutional neural network method to carry out research on the image recognition of small samples of tea tree diseases， and basically realized the effective discrimination of three kinds of easily mixed diseases in the case of small samples. Jiang et al.[5] proposed a ginger disease recognition system based on convolutional neural network， and developed a human-computer interaction interface using Qt， which showed the four diseases corresponding to ginger and their probabilities. Guo et al.[6] proposed a multi-scale recognition model suitable for mobile platforms， which reduced system memory requirements by 95.4%. Zhang et al.[7] proposed an apple leaf disease identification method based on IDCNNs， which could directly extract classification features from disease images and successfully distinguish three apple leaf diseases using Softmax classifier. Hu et al.[8] used median filtering combined with K-means clustering method to segment wheat powdery mildew， stripe rust and leaf rust， and then extracted the color and texture feature parameters. They designed a feature selection method based on the combination of the variance algorithm primary selection and the sequence floating forward selection search algorithm （SFFS） to select a subset of good features to realize the identification of three kinds of diseases in the leaves of wheat. Wang et al.[9] proposed a small multi-rotor UAV rice white head disease recognition system， which extracted Haar-like features using the UAV platform as the basis of image processing and recognition， and used Adaboost algorithm for white head training recognition， and its recognition rate reached 93.62%. Lu et al.[10] established an SVR prediction model for three diseases of lettuce based on the characteristics of hyperspectral and image fusion， and the recognition rate was 92.23%， but it is expensive. In this study， an extraction method combining color， morphology， texture and other characteristics was proposed to identify three kinds of diseases of black spot， anthracnose and leaf blight in O. japonicus leaves， so as to adapt to practical application.

Disease Image Processing

Image preprocessing

In July 2019， 60 images of three diseases （black spot， anthracnose， and leaf blight） were collected under natural light conditions at an O. japonicus planting base in Qingbaijiang area of Chengdu City in July 2019， of which 30 were used as training samples， and other 30 were the test samples. In order to reduce the amount of calculation and increase the operating speed of the system， the original images were uniformly tailored to a size of 400×300 pixels， and irrelevant redundant information was deleted at the same time. Then， the histogram equalization technology was used to enhance the tailored disease images， so that the detailed information of the images was clearer and the quality and recognizability of images were improved， which was convenient for people and computers to further analyze and process the images. In addition， there was inevitably some noise in the process of image acquisition and transmission[11]. Therefore， we continued to use the median filtering technology to further denoise the enhanced diseased images， so as to achieve a better smoothing effect （Fig. 1）.

Comparison of the effect of lesion segmentation

In order to further extract the characteristic information of disease spots， the bimodal method， the Otsu method and the K-means clustering algorithm were applied to segment the images of the three kinds of O. japonicus leaf disease. The effects are shown in Table 1. It can be seen from Table 1 that the K-means clustering segmentation algorithm could directly process color disease images， remove most of the background， and then perform mathematical morphological processing on them to obtain relatively complete disease spot images. Meanwhile， it avoided over-segmentation and incomplete segmentation.

Feature Extraction and Optimization

Disease image feature extraction is one of the key steps of digital image processing， which is related to the efficiency and recognition rate of classifiers[12]. Therefore， it is necessary to extract the features that distinguish the disease spots clearly with better robustness. Common disease spot feature extraction methods are divided into color feature， morphological feature and texture feature extraction based on different feature attributes.

Extraction of color features

Color is the most direct visual feature to describe disease spots and has less dependence on the direction， size， and viewing angle of images themselves， and has strong robustness[13-15]. Color moments are based on mathematical methods， and further describe the color distribution in images by calculating moment features， and these color feature information are mainly concentrated in low-order moments[11]. In other words， only the first-order moment， second-order moment， and third-order moment of disease spot images need to be calculated to describe the color information of diseased spot images. The calculation is shown in formula （1）.

μi=1N∑Nj=1Pij

σi=1N∑Nj=1Pij-μi21/2

ζi=1N∑Nj=1Pij-μi31/3（1）

Wherein μi， σi， and ζi are the first-order moment， second-order moment， and third-order moment of the image respectively; Pij is the jth color component of the ith pixel; and N is the number of pixels.

The first-order moment， second-order moment， and third-order moment of each color component of the disease spot images in the RGB and HSV color spaces were calculated respectively （Fig. 2）. It can be seen from Fig. 2 that there were no obvious differences in the second-order moments of the RGB color space， and other feature parameters had a certain degree of discrimination， so they could be used as color features to distinguish different diseases.

Extraction of shape features

The combination of disease spot shape and size measurement is also an important basis for distinguishing different disease spots[12]. Based on the disease spot images， the following geometric features were extracted： ① disease spot size A was extracted， which was expressed as the number of pixels with a gray value of 1 in binary images. This feature is greatly affected by size， distortion， zoom， and shooting conditions. ② Canny operator was used to extract the edges of the disease spot images and then the number of boundary pixels was counted， that is， the circumference of disease spot L. ③ Circularity is also called quasi-circularity， which represents the similarity between the shape of a target image in images and circle. It is defined as the ratio of 4π times the area of the area A to the square of its perimeter L， that is， 4πA/L2. ④ Rectangularity describes the fullness of a target in its minimum enclosing rectangle area. The rectangularity of a rectangle， a circle and an equilateral triangle is  π/4， and 0.5， respectively， and the rectangularity of other irregular shapes is between 0 and 1. In other words， the rectangularity value can be used to distinguish rectangles， circles， and irregular shapes. ⑤ The aspect ratio of the smallest bounding rectangle， which is not affected by the size and direction of the image， is an ideal geometric shape feature. Based on this， the five geometric features （area A， perimeter L， circularity C， rectangularity Ro， aspect ratio Ra） of disease spot images were calculated and the results were plotted as Fig. 3. The two features of area A and perimeter L were greatly affected by the image collection environment， but there were certain differences in circularity， rectangularity， and the aspect ratio of the minimum enclosing rectangle， which could be used as the shape features to distinguish different disease spots.

Tao YANG et al. Recognition of Ophiopogon japonicus Disease Based on Image Feature Fusion

Extraction of texture features

The texture feature of disease spot images refers to the slowly changing or periodically changing surface structure and organization of the disease spot surface， which reflects the roughness and granularity of disease spots. Different disease spots often show different textures[13]. Common texture feature description methods mainly include gray-level difference statistics method， autocorrelation function method， gray-level co-occurrence matrix method and spectral analysis method[14]. Based on the gray-level co-occurrence matrix， the energy， contrast， correlation and entropy values in the four directions of 0°， 45°， 90°， and 135° were calculated at the distance d=2 （Fig. 4）. It can be seen from Fig. 4 that only the contrast， correlation and entropy of disease spots in the 0° direction were significantly different， which could be used as texture features to distinguish different disease spots.

Feature selection and optimization

When extracting disease features， it is necessary to extract feature parameters that may help improve systems recognition rate as many as possible. In fact， there is some relevant redundant information， which will bring a huge amount of calculation and affect the performance of classifiers[15-16]. Based on this， we combined variance comparison method and principal component analysis （Principal Component Analysis， PCA） for feature selection and optimization. First of all， two outlier feature parameters of area A and perimeter L that had large variance and were greatly influenced by the outside world were removed. Secondly， the feature parameters with poor disease characterization ability such as the second moment of RGB color components were eliminated， and only the first and third moments of each color component of RGB， the color moment of each color component of HSV， circularity C， rectangularity Ro， aspect ratio Ra， contrast， correlation and entropy in the 0° direction， a total of 21 feature parameters were reserved and fused into disease image feature vectors. Then， PCA dimensionality reduction technology was used to reduce the feature vectors to 10 dimensions.

Disease Identification

Support vector machine （SVM） is an effective tool for solving pattern recognition and function estimation problems， especially the use of kernel functions to calculate in low-dimensional space， which better solves the "curse of dimensionality" problem， i.e.， the cumbersome and difficult calculations in high-dimensional space. It performs especially well in small-sample classification and recognition tasks[17-18]. Based on this， three SVM classifiers were designed. In this study， the sample to be tested x was input into these three classifiers for classification and recognition， and the three classification results were respectively denoted as fA（x）， fB（x） and fC（x）. If the sample belonged to the category， the return value would be 1; otherwise， the return value would be -1.

fq（x）= Belong to this category- Exclude from this category， q∈{A， B， C}（2）

However， this identification method may have a situation where fq（x） are all negative or multiple values are positive. In other words， the classifier fails to correctly give the unique recognition result. Therefore， it is necessary to reasonably set the penalty factor c and the width of the kernel function σ to avoid the failure in recognization and obtain a more ideal recognition rate[19-20]. In this study， 180 images of O. japonicus leaf diseases were selected. Through trial and error， when c=0.8 and σ=0.6， the recognition effect was better （Table 2）. Leaf blight had a better recognition effect due to its obvious characteristics and larger area of diseased spots; black spot took the second place; and anthracnose disease had a slightly lower recognition rate due to its color characteristic of the disease spots was similar to those of normal green leaves.

Conclusions

In this study， we extracted the color， shape， and texture feature information of disease spot images， and successfully identified three diseases of O. japonicus leaves using K-means clustering segmentation algorithm， PCA， SVM and other technologies. Compared with the convolutional neural network method， this method has the advantages of low cost， simple algorithm， high efficiency， etc.， and basically meets requirements of application. This study can provide a theoretical basis for the prevention of O. japonicus diseases， and promote the development of the O. japonicus industry toward informatization.

References

[1] MU FM， LIU S， LI M. Current Situation and development prospect of industry of Ophiopogon japonicus in Sichuan[J]. Northern Horticulture， 2019（17）： 151-157. （in Chinese）

[2] LIU CC， YANG T， MA JJ， et al. Identification system for leaf diseases of Ophiopogon japonicus based on PCA-SVM[J]. Journal of Chinese Agricultural Mechanization， 2019， 40（8）： 132-136. （in Chinese）

[3] WU N. Research on key techniques of crop leaf diseases recognition based on image retrieval[D]. Hefei： University of Science and Technology of China， 2018. （in Chinese）

[4] SUN YY， JIANG CH， DONG W， et al. Image recognition of tea plant disease based on convolutional neural net-work and small samples[J]. Jiangsu Journal of Agricultural Sciences， 2019， 35（1）： 48-55. （in Chinese）

[5] JIANG FQ， LI Y， YU DW， et al. Design and experiment of tobacco leaf grade recognition system based on Caffe[J]. Journal of Chinese Agricultural Mechanization， 2019， 40（1）： 126-131. （in Chinese）

[6] GUO XQ， FAN TJ， SHU X. Tomato leaf diseases recognition based on improved Multi-Scale AlexNet[J]. Transactions of the Chinese Society of Agricultural Engineering， 2019， 35（13）： 162-169. （in Chinese）

[7] ZHANG SW， ZHANG QQ， LI P. Apple disease identification based on improved deep convolutional neural network[J]. Journal of Forestry Engineering， 2019， 4（4）： 107-112. （in Chinese）

[8] HU WW， ZHANG W， LIU LZ. Identification of wheat leaf diseases based on variance-SFFS algorithm[J]. Journal of Hunan Agricultural University： Natural Science， 2018， 44（2）： 225-228. （in Chinese）

[9] WANG Z， CHU GK， ZHANG HJ， et al. Identification of diseased empty rice panicles based on Haar-like feature of UAV optical image[J]. Transactions of the Chinese Society of Agricultural Engineering， 2018， 34（20）： 73-82. （in Chinese）

[10] LU B， SUN J， MAO HP， et al. Disease recognition of lettuce with feature fusion based on hyperspectrum and image[J]. Jiangsu Journal of Agricultural Sciences， 2018， 34（6）： 1254-1259. （in Chinese）

[11] QIAO H， FENG Q， ZHANG R， et al. Dynamic monitoring of grape leaf disease based on sequential images tracking[J]. Transactions of the Chinese Society of Agricultural Engineering， 2018， 34（17）： 167-175. （in Chinese）

[12] WANG YX， ZHANG Y， YANG CY， et al. Advances in new nondestructive detection and identification techniques of crop diseases based on deep learning[J]. Acta Agriculturae Zhejiangensis， 2019， 31（4）： 669-676. （in Chinese）

[13] LI XX， ZHU CG， BAI XB， et al. Recognition method of cucumber leaves diseases based on visual spectrum and support vector machine[J]. Spectroscopy and Spectral Analysis， 2019， 39（7）： 2250-2256. （in Chinese）

[14] GUO XQ， FAN TJ， SHU X. Recognition of tomato leaf disease based on image fusion feature[J]. Journal of Hunan Agricultural University： Natural Sciences， 2019， 45（2）： 212-217， 224. （in Chinese）

[15] CHEN LL， JIANG DQ， CAI YJ， et al. Monitoring and identification of the taro disease based on the computer vision[J]. Journal of Agricultural Mechanization Research， 2020， 42（6）： 224-229. （in Chinese）

[16] WU YR， LI JH. Automatic recognition of cucumber disease based on joint sparse model[J]. Journal of Hunan Agricultural University： Natural Sciences， 2019， 45（4）： 444-448. （in Chinese）

[17] ZHANG SW， XIE ZQ， ZHANG QQ. Application research on convolutional neural network for cucumber leaf disease recognition[J]. Jiangsu Journal of Agricultural Sciences， 2018， 34（1）： 56-61. （in Chinese）

[18] XIA YQ， WANG B， ZHI J， et al. Identification of wheat leaf disease based on random forest method[J]. Journal of Graphics， 2018（1）： 57-62. （in Chinese）

[19] PENG YK， ZHAO F， LI L， et al. Discrimination of heat-damaged tomato seeds based on near infrared spectroscopy and PCA-SVM method[J]. Transactions of the Chinese Society of Agricultural Engineering， 2018， 34（5）： 159-165. （in Chinese）

[20] HUO YQ， ZHANG C， LI YH， et al. Nondestructive detection for kiwifruit based on the hyperspectral technology and machine learning[J]. Journal of Chinese Agricultural Mechanization， 2019， 40（4）： 71-77. （in Chinese）