An online identification method based on internal freezing damage of citrus

Abstract

The invention discloses an online recognition method based on internal frost damage of citrus, and relates to the technical field of agricultural production. Citrus fruit images and real-time environmental information are collected for preprocessing, and image features of citrus fruits with internal frost damage in an image database are extracted as input for pretraining to obtain a first neural network model, and feature parameters in reference environmental information are extracted as input for pretraining to obtain a second neural network model, and image feature vectors in the image to be identified and environmental feature vectors of real-time environmental information are extracted through the model, and the image feature vectors and environmental feature vectors are fused by a data fusion algorithm, and whether citrus fruits in the image to be identified are frost damaged is determined according to the feature vector fusion result, and whether frost damage occurs to the internal fruits of citrus is determined based on the image recognition and data fusion method, and environmental factors are added for comprehensive analysis to determine whether frost damage occurs to the internal fruits of citrus, so as to improve the accuracy of the recognition result.

Description

Technical Field

[0001] The invention relates to the technical field of agricultural production, and in particular to an online identification method based on internal freezing damage of citrus. Background Art

[0002] With the development and progress of science and technology, more and more cutting-edge technologies are being applied to agricultural production. At present, the non-destructive detection technology for citrus fruits mainly involves infrared spectrum recognition, X-ray recognition and image recognition based on machine learning. However, the above detection methods are all based on the recognition and analysis of the characteristics of citrus images to judge the quality of the internal fruit of citrus. In the detection and recognition process, the influence of environmental factors is not considered, resulting in a low accuracy rate in determining whether frost damage occurs inside citrus fruits. To this end, we propose an online recognition method based on frost damage inside citrus. Summary of the invention

[0003] The main purpose of the present invention is to provide an online recognition method based on internal frost damage of citrus, which collects citrus fruit images and real-time environmental information for preprocessing, extracts the image features of citrus fruits with internal frost damage in the image database as input for pretraining to obtain a first neural network model, extracts the feature parameters in the reference environmental information as input for pretraining to obtain a second neural network model, and extracts the image feature vector in the image to be identified and the environmental feature vector of the real-time environmental information through the model, fuses the image feature vector with the environmental feature vector using a data fusion algorithm, and determines whether the citrus fruit in the image to be identified has suffered frost damage according to the feature vector fusion result. The problem in the background technology can be effectively solved.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] An online identification method based on internal freezing damage of citrus, comprising:

[0006] Collecting several groups of citrus fruit images, screening the collected images and performing preprocessing, including graying, geometric transformation, and image enhancement, to obtain images to be identified;

[0007] Real-time environmental information is collected by an environmental sensor during the process of capturing images of citrus fruits. The real-time environmental information and the reference environmental information include at least one of a temperature value, a humidity value, a light intensity value, a temperature change per unit time, a humidity change, and a light change. The acquired real-time environmental information is preprocessed according to the Grubbs criterion, a data critical value of the Grubbs criterion is set, and coarse data in the real-time environmental information that is greater than the data critical value is eliminated, wherein:

[0008] The statistical value expression of Grubbs criterion is:

[0009]

[0010] Among them, T i for i The statistical value of the Grubbs criterion for each sensor; For the i The average value of sensor data; x ij For the i The data of the jth time node of a sensor; n is the total number of data; σ i For the i The standard deviation of the sensor data,

[0011] The images of citrus fruits with internal frost damage are obtained, an image database is constructed, and the image features of citrus fruits with internal frost damage in the image database are extracted as input for pre-training to obtain a first neural network model. The image features include fruit shape parameters, color parameters and texture parameters, wherein:

[0012] The fruit shape parameters include shape characteristics and contour characteristics, specifically at least one of a single fruit area value, a fruit outer contour perimeter value, a major diameter value, a minor diameter value, a major diameter to minor diameter ratio, and roundness;

[0013] The color parameter includes at least one of an HSV value, an HSI value, and an RGB value;

[0014] The texture parameters include at least one of the brightness level of the texture, the contrast of the texture, the complexity of the texture, the directionality of the texture, the texture density, and the connectivity of the texture. The first neural network model adjusts the number of hidden layers of the model according to the error between the training result and the actual alarm result until the accuracy is not less than the first expected value, wherein the calculation formula of the first expected value is,

[0015]

[0016] Where E(Y)1 represents the first expected value; N1 represents the number of input samples of the first neural network model; f(X i ) represents the output function of the first neural network model; X i represents the i-th output sample of the first neural network model, and extracts the image feature vector in the image to be recognized according to the constructed first neural network model;

[0017] Reference environment information when collecting images of citrus fruits with internal frost damage is obtained, and feature parameters in the reference environment information are extracted as input for pre-training to obtain a second neural network model. The second neural network model adjusts the number of hidden layers of the model according to the error between the training result and the actual alarm result until the accuracy is not less than the second expected value. The calculation formula of the second expected value is,

[0018]

[0019] Where E(Y)2 represents the second expected value; N2 represents the number of input samples of the second neural network model; f(X j ) represents the output function of the second neural network model; X j represents the j-th output sample of the second neural network model, and extracts the environmental feature vector of the real-time environmental information according to the constructed second neural network model;

[0020] The data fusion algorithm is used to fuse the image feature vector with the environmental feature vector to obtain a fused feature vector. The fusion result of the feature vector is used to determine whether the citrus fruits in the image to be identified are frozen.

[0021] The present invention has the following beneficial effects:

[0022] (1) Compared with the prior art, the recognition method proposed in the technical solution of the present invention collects citrus fruit images, pre-processes the collected images after screening, obtains the image to be recognized, collects real-time environmental information in the process of shooting the citrus fruit images through an environmental sensor, pre-processes the obtained real-time environmental information according to the Grubbs criterion, sets a data critical value of the Grubbs criterion, removes coarse data in the real-time environmental information that is greater than the data critical value, obtains images of citrus fruits with internal frost damage to build an image database, extracts image features of citrus fruits with internal frost damage from the image database as input for pre-training to obtain a first neural network model, and extracts image features in the image to be recognized according to the constructed first neural network model. The image is imaged using a sensor to identify citrus fruits with internal frost damage. ... BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention is a flow chart of an online identification method based on internal freezing damage of citrus. DETAILED DESCRIPTION

[0024] The present invention will be further described below in conjunction with specific implementation methods, wherein the accompanying drawings are only used for exemplary descriptions and represent only schematic diagrams rather than actual drawings, and should not be understood as limiting the present invention. In order to better illustrate the specific implementation methods of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product.

[0025] Example 1

[0026] like Figure 1 A flow chart of an online identification method based on internal frost damage of citrus is shown.

[0027] The specific implementation steps of the identification method proposed by the technical solution of the present invention are as follows:

[0028] Step 1, collecting a number of groups of citrus fruit images, screening the collected images and performing preprocessing, including graying, geometric transformation, and image enhancement, to obtain images to be identified;

[0029] It should be noted that during the collection process, citrus canopy images should be taken under natural light as much as possible, and fruits with shadows under the canopy should be collected. To facilitate the training of the subsequent neural network model, citrus images can be taken at different distances during collection. The collection environment should preferably include cloudy days and sunny days, covering a variety of lighting conditions such as front light and back light. After obtaining the image, the Matlab tool is used to amplify the original data set. Data enhancement is performed by using random flipping, local cropping, length and width scaling, contrast-limited adaptive histogram equalization, median filtering, adding Gaussian noise, adding gamma noise, etc. to obtain the image to be identified;

[0030] Furthermore, several groups of cameras can be set up in the citrus orchard, and the shooting angles of the cameras can be adjusted to shoot the crowns of the citrus trees. The captured images can be sent via communication equipment to achieve remote detection.

[0031] Step 2, collecting real-time environmental information in the process of shooting citrus fruit images through environmental sensors. It should be noted that, as in step 1, several groups of cameras can be set up in the citrus orchard to shoot the citrus tree crowns. Similarly, sensors can be set up in the orchard to obtain implementation environmental information. The data of the sensors can be sent through communication equipment to achieve remote detection. The real-time environmental information and reference environmental information include at least one of temperature value, humidity value, light intensity value, temperature change per unit time, humidity change, and light change. The acquired real-time environmental information is preprocessed according to the Grubbs criterion, and the data critical value of the Grubbs criterion is set to eliminate coarse data in the real-time environmental information that is greater than the data critical value.

[0032] The statistical value expression of Grubbs criterion is:

[0033]

[0034] Among them, T i for i The statistical value of the Grubbs criterion for each sensor; For the i The average value of sensor data; x ij For the i The data of the jth time node of a sensor; n is the total number of data; σ i For the i The standard deviation of the sensor data,

[0035] Step 3: Acquire images of citrus fruits with internal frost damage. The acquisition method is the same as the operation of collecting citrus fruit images in step 1, and will not be repeated here. An image database is constructed based on the acquired images of citrus fruits with internal frost damage. The image features of citrus fruits with internal frost damage in the image database are extracted as input for pre-training to obtain a first neural network model. The image features include fruit shape parameters, color parameters and texture parameters, wherein:

[0036] The fruit shape parameters include shape characteristics and contour characteristics, specifically at least one of a single fruit area value, a fruit outer contour perimeter value, a major diameter value, a minor diameter value, a major diameter to minor diameter ratio, and roundness;

[0037] The color parameter includes at least one of an HSV value, an HSI value, and an RGB value;

[0038] The texture parameters include at least one of the brightness level of the texture, the contrast of the texture, the complexity of the texture, the directionality of the texture, the texture density, and the connectivity of the texture. The first neural network model adjusts the number of hidden layers of the model according to the error between the training result and the actual alarm result until the accuracy is not less than the first expected value, wherein the calculation formula of the first expected value is,

[0039]

[0040] Where E(Y)1 represents the first expected value; N1 represents the number of input samples of the first neural network model; f(X i ) represents the output function of the first neural network model; X i represents the i-th output sample of the first neural network model, and extracts the image feature vector in the image to be recognized according to the constructed first neural network model;

[0041] Step 4: obtain reference environmental information when collecting images of citrus fruits with internal frost damage, extract feature parameters in the reference environmental information as input for pre-training to obtain a second neural network model, and adjust the number of hidden layers of the second neural network model according to the error between the training result and the actual alarm result until the accuracy is not less than the second expected value. The calculation formula of the second expected value is:

[0042]

[0043] Where E(Y)2 represents the second expected value; N2 represents the number of input samples of the second neural network model; f(X j ) represents the output function of the second neural network model; X j represents the j-th output sample of the second neural network model, and extracts the environmental feature vector of the real-time environmental information according to the constructed second neural network model;

[0044] Step 5: Using a data fusion algorithm to fuse the image feature vector with the environmental feature vector to obtain a fused feature vector, and judging whether the citrus fruit in the image to be identified is frozen or not according to the feature vector fusion result;

[0045] It should be noted that in this step, the YOLO data fusion model in the prior art can be selected to adopt the add parallel strategy to fuse two groups of feature vectors, namely the image feature vector and the environmental feature vector, to obtain a group of fused feature vectors. This group of vector features combines the environmental feature vector extracted by the second neural network model to extract real-time environmental information and the image feature vector extracted by the first neural network model to be identified. Environmental factors are added to the existing image recognition technology to make the judgment strategy more in line with the actual situation, which is conducive to improving the accuracy of identifying internal frost damage in citrus.

[0046] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An online identification method based on internal freezing damage of citrus, characterized in that: include, Collecting several groups of citrus fruit images, screening the collected images and performing preprocessing, including graying, geometric transformation, and image enhancement, to obtain images to be identified; The real-time environmental information is collected by the environmental sensor in the process of shooting the citrus fruit image, the real-time environmental information is pre-processed according to the Grubbs criterion, the data critical value of the Grubbs criterion is set, and the coarse data in the real-time environmental information that is greater than the data critical value is eliminated; Acquire images of citrus fruits with internal frost damage, construct an image database, extract image features of citrus fruits with internal frost damage in the image database as input for pre-training to obtain a first neural network model, and extract image feature vectors in the image to be identified based on the constructed first neural network model; Acquire reference environmental information when collecting images of citrus fruits with internal frost damage, extract characteristic parameters in the reference environmental information as input for pre-training to obtain a second neural network model, and extract environmental feature vectors of real-time environmental information according to the constructed second neural network model; The data fusion algorithm is used to fuse the image feature vector with the environmental feature vector, and the fusion result of the feature vector is used to determine whether the citrus fruits in the image to be identified are frozen.

2. The online identification method based on internal freezing damage of citrus according to claim 1, characterized in that: The statistical value expression of Grubbs criterion is: Among them, T i for i The statistical value of the Grubbs criterion for each sensor; For the i The average value of sensor data; x ij For the i The data of the jth time node of a sensor; n is the total number of data; σ i For the i The standard deviation of the sensor data, 3. The online identification method based on internal freezing damage of citrus according to claim 1, characterized in that: The real-time environment information and the reference environment information include at least one of a temperature value, a humidity value, a light intensity value, a temperature change per unit time, a humidity change, and a light change.

4. The online identification method based on internal freezing damage of citrus according to claim 1, characterized in that: The first neural network model adjusts the number of hidden layers of the model according to the error between the training result and the actual alarm result until the accuracy is not less than the first expected value, wherein the calculation formula of the first expected value is: Where E(Y)1 represents the first expected value; N1 represents the number of input samples of the first neural network model; f(X i ) represents the output function of the first neural network model; X i Represents the i-th output sample of the first neural network model.

5. The online identification method based on internal freezing damage of citrus according to claim 1, characterized in that: The second neural network model adjusts the number of hidden layers of the model according to the error between the training result and the actual alarm result until the accuracy is not less than the second expected value. The calculation formula of the second expected value is: Where E(Y)2 represents the second expected value; N2 represents the number of input samples of the second neural network model; f(X j ) represents the output function of the second neural network model; X j Represents the j-th output sample of the second neural network model.

6. The online identification method based on internal freezing damage of citrus according to claim 1, characterized in that: The method for determining rough data is: The data critical value is T(n0,a), where n0 represents the number of times the sensor node measures the data, and a represents the excellent level of the data, which is between 0.01 and 0.

05. When the statistical value of the Grubbs criterion of any sensor node data is T i When it is greater than the critical value, the data of the sensor node is determined to be coarse data.

7. The online identification method based on internal freezing damage of citrus according to claim 1, characterized in that: Image features include fruit shape parameters, color parameters and texture parameters, among which: The fruit shape parameters include shape characteristics and contour characteristics, specifically at least one of a single fruit area value, a fruit outer contour perimeter value, a major diameter value, a minor diameter value, a major diameter to minor diameter ratio, and roundness; The color parameter includes at least one of an HSV value, an HSI value, and an RGB value; The texture parameter includes at least one of a brightness level of the texture, a contrast of the texture, a complexity of the texture, a directionality of the texture, a density of the texture, and a connectivity of the texture.

Images